JP7575007B2 - Hand load estimation device, robot control system, and robot system - Google Patents

Description

The present disclosure relates to a hand load estimation device that estimates parameters regarding the load applied to a robot's hand, a robot control system, and a robot system.

When controlling industrial and other robots, it is necessary to appropriately set parameters such as the mass or center of gravity position of the load applied to the hand that is the object of operation so that the robot can perform high-speed operations and highly accurate positioning. Hereinafter, parameters related to the load applied to the robot's hand are referred to as hand load parameters.

Patent Document 1 discloses a method for identifying hand load parameters for a multi-joint robot based on the state quantities of each joint when the robot is operated. In the method of Patent Document 1, hand load parameters are identified from state quantities such as the position, speed, acceleration, and current required for operation of each joint of the robot, based on a linearized equation for the hand load parameters and the drive torque shown in the robot model. According to the method of Patent Document 1, hand load parameters can be estimated without limiting the type of load or the robot's operation pattern.

Patent No. 4137673

In the method of Patent Document 1, the hand load parameters are estimated based only on the data acquired for load estimation. For this reason, in the case of the method of Patent Document 1, the estimated hand load parameters may have errors from the actual hand load parameters due to the influence of modeling errors in the robot model or the influence of characteristics not included in the robot model. For this reason, the conventional technology of Patent Document 1 has the problem that it is difficult to obtain highly accurate estimates of the hand load parameters.

The present disclosure has been made in consideration of the above, and aims to provide a hand load estimation device that can obtain highly accurate estimation results of hand load parameters.

In order to solve the above-mentioned problems and achieve the object, the hand load estimation device according to the present disclosure is a hand load estimation device that estimates a hand load parameter that is a parameter regarding the load applied to the hand of a robot. The hand load estimation device according to the present disclosure includes an inverse dynamics model unit that outputs a nominal torque value that is a torque calculated based on an inverse dynamics model by inputting motion data, which is at least one of data on the position, speed, and acceleration of a joint of the robot, into the inverse dynamics model; a learning unit that learns a hand load parameter from which the influence of modeling errors of the inverse dynamics model and the influence of characteristics of the robot that are not included in the inverse dynamics model are removed, based on learning data including the motion data and the difference between the actual torque and the nominal torque of the joint; and an inference unit that receives inference data including the motion data and the difference between the actual torque and the nominal torque, and infers the hand load parameter based on the learning result by the learning unit.

The hand load estimation device disclosed herein has the advantage of being able to obtain highly accurate estimation results of hand load parameters.

FIG. 1 is a diagram showing a configuration of a hand load estimation device according to a first embodiment; FIG. 1 is a diagram showing a configuration of a hand load estimation device according to a first embodiment when the hand load estimation device performs supervised learning of hand load parameters using an RNN; FIG. 1 is a conceptual diagram showing a configuration of a neural network used for learning in the first embodiment. FIG. 4 shows an example of an LSTM unit constituting the neural network shown in FIG. 3. FIG. 1 is a diagram showing a configuration of a hand load estimating device according to a first embodiment in a case where the hand load estimating device uses a torque value detected by a torque sensor as an actual torque value; FIG. 1 is a diagram showing a configuration of a hand load estimation device according to a second embodiment; FIG. 11 is a diagram for explaining GPRs used in learning in the second embodiment. FIG. 13 is a diagram for explaining a histogram filter used in inference in the second embodiment. FIG. 13 is a diagram showing a configuration of a hand load estimation device according to a third embodiment. FIG. 13 is a diagram showing a configuration of a hand load estimation device according to a fourth embodiment. FIG. 13 is a diagram showing a configuration of a hand load estimation device according to a fifth embodiment. FIG. 13 is a diagram showing a configuration of a hand load estimation device according to a sixth embodiment. FIG. 13 is a diagram showing a configuration of a hand load estimation device according to a seventh embodiment. FIG. 23 is a conceptual diagram for explaining updating of a probability distribution by Bayesian linear regression in the seventh embodiment. FIG. 13 is a diagram showing a configuration of a robot system according to an eighth embodiment. FIG. 1 is a diagram showing a first example of a hardware configuration for realizing a hand load estimation device according to the first to seventh embodiments. FIG. 13 is a diagram showing a second example of the configuration of hardware for realizing the hand load estimation device according to the first to seventh embodiments.

The hand load estimation device, robot control system, and robot system according to the embodiments are described in detail below with reference to the drawings.

Embodiment 1.
FIG. 1 is a diagram showing a configuration of a hand load estimation device 1 according to a first embodiment. The hand load estimation device 1 estimates a hand load parameter, which is a parameter regarding a load applied to a hand of a robot. In the first embodiment, the robot is an industrial robot. The robot has a plurality of joints that operate a movable part of the robot. The hand is a tip part of the movable part. A load is attached to the hand. An example of the load is a detachable part such as a hand or a tool.

The hand load estimation device 1 has an input unit 11, an inverse dynamics model unit 12, a learning unit 13, a learning result storage unit 14, and an inference unit 15. Data on the internal state quantities of the robot are input to the input unit 11 for each control cycle of the robot. In this way, the hand load estimation device 1 obtains time series data on the internal state quantities of the robot. The state quantity data may be input for each integer multiple of the control cycle.

The state quantity data includes position data and current data. The position data is data that represents the position of the joint. Here, the position data is a feedback value of the joint position. The position of the joint is the position in the rotational direction of the joint, and is represented by, for example, a rotation angle. The current data is data that represents the current flowing through the motor that rotates the joint. Here, the current data is a feedback value of the current flowing through the motor. Position data, which is the position feedback (FB) value of each joint, and current data, which is the current feedback (FB) value of each joint are input to the input unit 11.

When the learning unit 13 is learning, the input unit 11 acquires time series data of the position FB value and time series data of the current FB value as input data during learning. When the inference unit 15 is inferring, the input unit 11 acquires time series data of the position FB value and time series data of the current FB value as input data during inference. Figure 1 shows an image of the input data of the input unit 11 during learning and an image of the input data of the input unit 11 during inference.

The position θ, velocity θ' and acceleration θ" of each joint are calculated from the position data input to the input unit 11. Means for calculating the position θ, velocity θ' and acceleration θ" from the position data are omitted from the illustration. Movement data, which is data on the position θ, velocity θ' and acceleration θ" of each joint calculated from the position data of each joint, is input to the inverse dynamics model unit 12. The inverse dynamics model is a model that calculates torque by solving an equation of motion. The position θ, velocity θ' and acceleration θ" of each joint are input to the inverse dynamics model, and the inverse dynamics model unit 12 outputs a nominal torque τ _mdl of each joint. The nominal torque τ _mdl is a torque calculated from the equation of motion for the joint. The inverse dynamics model is built into the hand load estimation device 1 in advance.

When calculating the equation of motion, the inverse dynamics model unit 12 refers to pre-set nominal parameters. Parameters such as the mass, center of gravity position, or inertia tensor of each link of the robot are set as the nominal parameters. Values at the time of designing the robot can be used for the nominal parameters. Alternatively, values previously identified by the robot from which position data and current data are acquired can be used for the nominal parameters.

During learning by the learning unit 13, the inverse dynamics model unit 12 calculates the nominal torque τ _{mdl based on the position θ, velocity θ', and acceleration θ" obtained from the input data during learning. During inference by the inference unit 15, the inverse dynamics model unit 12 calculates the nominal torque τ mdl based} on the position θ, velocity θ', and acceleration θ" obtained from the input data during inference. Note that in FIG. 1, for the convenience of separately showing the data flow during learning and the data flow during inference, the inverse dynamics model unit 12 is shown as being divided into two.

The actual torque τ _real is obtained by multiplying the current data input to the input unit 11 by a torque constant. That is, the hand load estimation device 1 uses the value obtained by multiplying the current data by the torque constant as the τ _real value of the actual torque. The hand load estimation device 1 calculates the difference τ _real -τ _mdl between the actual torque τ _real and the nominal torque τ _mdl output from the inverse dynamics model unit 12.

The learning unit 13 receives as learning data the position θ, velocity θ', and acceleration θ'' of each joint, which is movement data obtained from the input data during learning, and the difference τ _real - τ _mdl of each joint based on the input data during learning. The learning data for each joint is input to the learning unit 13. Using the learning data, the learning unit 13 learns hand load parameters from which the influence of modeling errors in the inverse dynamics model and the influence of characteristics of the robot that are not included in the inverse dynamics model have been removed. Hereinafter, characteristics of the robot that are not included in the inverse dynamics model are referred to as outside-model characteristics. The learning unit 13 stores the learning results in the learning result storage unit 14.

The inference unit 15 receives, as inference data, the position θ, velocity θ', and acceleration θ'' of each joint, which is motion data obtained from the input data at the time of inference, and the difference τ _real -τ _mdl of each joint based on the input data at the time of inference. The inference data for each joint is input to the inference unit 15. The inference unit 15 also reads out the learning result by the learning unit 13 from the learning result storage unit 14. The inference unit 15 estimates hand load parameters based on the learning result by the learning unit 13. In the configuration shown in FIG. 1, the hand load estimation device 1 can apply any method as the learning method in the learning unit 13. The hand load parameter is at least one of the parameters of the mass of the load, the center of gravity position of the load, and the inertia tensor of the load. In addition, the motion data input to each of the inverse dynamics model unit 12, the learning unit 13 and the inference unit 15 is not limited to data regarding all of the position θ, velocity θ' and acceleration θ'', but may be data regarding at least one of the position θ, velocity θ' and acceleration θ''.

The inference unit 15 uses, as inference data, the motion data based on the position FB value immediately after the robot starts to move, and the difference τ _real -τ _mdl based on the current FB value immediately after the robot starts to move. This allows the hand load estimation device 1 to estimate the hand load parameter at an early point in time after the robot starts to move. Also, it becomes possible to reflect the estimation result of the hand load parameter in the command value for controlling each joint at an early point in time after the robot starts to move.

Next, a specific example of the learning method will be described. Here, a case where supervised learning of hand load parameters is performed using an RNN (Recurrent Neural Network) will be described. Figure 2 is a diagram showing the configuration of the hand load estimation device 1 according to the first embodiment when the hand load estimation device 1 performs supervised learning of hand load parameters using an RNN.

The learning unit 13a of the hand load estimation device 1 shown in FIG. 2 receives as learning data the position θ, velocity θ', and acceleration θ'' of each joint, and the difference τ _real -τ _mdl of each joint based on the input data during learning. A teacher signal is also input to the learning unit 13a. The teacher signal is a correct answer label of the hand load parameter. The correct answer label is a hand load parameter that does not include errors due to modeling errors of the inverse dynamics model and out-of-model characteristics. The learning unit 13a constructs a neural network in which the output for the input of the position θ, velocity θ', acceleration θ'' and the difference τ _real -τ _mdl matches the correct answer label, and stores the learning result in the learning result storage unit 14a.

Examples of RNNs include Elman/Jordan net, LSTM (Long Short Term Memory), GRU (Gated Recurrent Unit), etc., and any of them may be used. In the first embodiment, the case where LSTM is used will be explained as an example.

Figure 3 is a conceptual diagram showing the configuration of a neural network used for learning in embodiment 1. The neural network includes an input layer 21, an intermediate layer 22, and an output layer 23. The input layer 21 includes neurons 24, the number of which corresponds to the number of elements of the input data, i.e., the number of inputs. In Figure 3, only the topmost neuron 24 in the input layer 21 is labeled, but circles in the input layer 21 with no label are also neurons 24. In Figure 3, the number of neurons 24 is three, but the number of neurons 24 is not limited to three.

The intermediate layer 22 includes a plurality of neurons 25. Each neuron 25 is an LSTM unit. In FIG. 3, only the topmost neuron 25 in the intermediate layer 22 is labeled, but double circles without a label in the intermediate layer 22 are also neurons 25. In FIG. 3, the intermediate layer is one layer, but the intermediate layer may be two or more layers. In FIG. 3, the number of neurons 25 is five, but the number of neurons 25 is not limited to five.

The output layer 23 has neurons 26, the number of which corresponds to the number of data elements to be output, i.e., the number of outputs. In FIG. 3, only the topmost neuron 26 in the output layer 23 is labeled, but circles in the output layer 23 with no label are also neurons 26. In FIG. 3, the number of neurons 26 is three, but the number of neurons 26 is not limited to three.

The input layer 21 outputs the input data to the intermediate layer 22. In detail, each neuron 24 outputs the input data to each neuron 25 in the intermediate layer 22. Each neuron 24 multiplies the input value by a weight and then outputs it to each neuron 25 in the intermediate layer 22. Each neuron 25 in the intermediate layer 22 performs LSTM processing and outputs the processing result to each neuron 26 in the output layer 23. Each neuron 26 in the output layer 23 adds the data output from each neuron 25 in the intermediate layer 22 and outputs the addition result. Each neuron 26 in the output layer 23 multiplies the data output from each neuron 25 in the intermediate layer 22 by a weight and adds the results of multiplication by the weights.

In an RNN, if a vector whose elements are input values is _xt and a vector whose elements are output values of the intermediate layer 22 is _ht , the vector _ht can be expressed by the following formula (1). t is a time discretized in units of a control period and is an integer indicating the ordinal control period. If the dimension of vector _xt is M and the dimension of vector _ht is N, W is an N×M dimensional linear transformation matrix, R is an N×N dimensional linear transformation matrix, and b is a bias vector. g(.) indicates an activation function.

^Let _xt1 , _xt2 , and _xt3 be input values that are elements of vector _xt ^. _Let ^ht1 , ^ht2 , _ht3 , _ht4 , _{and ht5} be output values that are ^elements of vector _ht . Let ^yt1 , _yt2 , and ^yt3 be output values that are elements of vector _yt . ^Let vector _yt be ^a _vector whose elements are the output ^values of _the output ^layer 23.

Figure 4 is a diagram showing an example of an LSTM unit that constitutes the neural network shown in Figure 3. In the configuration example shown in Figure 4, the LSTM unit performs the calculations shown in the following equations (2)-(7).

_Wf , _Wi , _Wz , and _Wo are linear transformation matrices of dimensions NxM, _Rf , _Ri , _Rz , and _Ro are linear transformation matrices of dimensions NxN, and _bf , _bi , _bz , and _bo are bias vectors. These are determined by learning by the learning unit 13a. σ(·) is a sigmoid function used as an activation function. tanh(·) is a hyperbolic tangent function used as an activation function.

In the first embodiment, the input layer 21 receives data on the position θ, velocity θ', acceleration θ'', and difference τ _real -τ _mdl . The vector x _t is a vector whose elements are the values of the position θ, velocity θ', acceleration θ'', and difference τ _real -τ _mdl . In the case of a six-axis vertical articulated robot, the dimension of the vector x _t is 24. The input data is not limited to data on all joints of the robot, and may be data on some of the joints. The input data may be data on only those joints that are in operation out of all joints of the robot. The learning unit 13a may reduce the dimension of the vector x _t by inputting data on only some of the joints.

In the first embodiment, the output layer 23 outputs values of hand load parameters estimated from data of the position θ, velocity θ', acceleration θ'', and difference τ _real -τ _mdl . The vector y _t is a vector whose elements are the values of the hand load parameters estimated from data at time t. The hand load parameters are parameters such as the mass of the load, the XYZ coordinates indicating the center of gravity of the load, the diagonal components I _xx , I _yy , and I _zz of the inertia tensor, and the off-diagonal components I _xy , I _yz , and I _zx of the inertia tensor. When all of these parameters are estimated, the dimension of the vector y _t is 10. The learning unit 13a may reduce the dimension of the vector y _t by outputting data of only some of these parameters.

The learning unit 13a applies an LSTM unit to the intermediate layer 22 to construct a neural network capable of extracting features that depend on past values of time-series data. The learning unit 13a also adjusts the weighting coefficients of each neuron 24, 25, 26 so that the output for the input of learning data matches the correct value of the hand load parameter. This allows the learning unit 13a to generate a trained model of an RNN that can absorb the effects of modeling errors and characteristics outside the model. The learning result storage unit 14a stores the trained model, which is the result of learning.

The inference unit 15a reads out the trained model of the RNN from the learning result storage unit 14a. In the inference unit 15a, each piece of data for inference, namely, the position θ, the velocity θ', the acceleration θ'' and the difference τ _real -τ _mdl, is input to the trained model. The inference unit 15a outputs the output of the trained model in response to the input as an estimation result of the hand load parameter.

According to the first embodiment, the hand load estimation device 1 learns hand load parameters from which the influence of modeling errors in the inverse dynamics model and the influence of characteristics outside the model have been removed, and infers the hand load parameters based on the learning results. The hand load estimation device 1 is capable of estimating hand load parameters from which errors due to the influence of modeling errors and the influence of characteristics outside the model have been reduced, thereby enabling highly accurate estimation of the hand load parameters immediately after the robot starts to operate. This provides the effect that the hand load estimation device 1 can obtain highly accurate estimation results of the hand load parameters.

The position data is a feedback value of the position of each joint, but is not limited to this. The position data may be a command value for controlling each joint. The hand load estimation device 1 obtains the command value from the robot control device.

The hand load estimation device 1 uses the value obtained by multiplying the current data by the torque constant as the value of the real torque τ _real , but is not limited to this. The hand load estimation device 1 may use the value of the torque detected by a torque sensor provided in the robot as the value of the real torque τ _real .

Figure 5 is a diagram showing the configuration of the hand load estimation device 1 in the case where the hand load estimation device 1 according to the first embodiment uses the torque value detected by the torque sensor 27 as the actual torque value. The torque sensor 27 detects the torque actually applied to the robot's hand and outputs the detected value. In the hand load estimation device 1 shown in Figure 5, instead of inputting the current data of the hand load estimation device 1 shown in Figure 1 and multiplying it by the torque constant, the detected value from the torque sensor 27 is input. "Torque output" in Figure 5 represents the detected value output from the torque sensor 27.

During learning by the learning unit 13, the input unit 11 acquires a detected torque value as input data during learning. During inference by the inference unit 15, the input unit 11 acquires a detected torque value as input data during inference. The hand load estimation device 1 calculates a difference τ _real - _{τ mdl} between the actual torque τ _real , which is the detected torque value, and the nominal torque τ _mdl output from the inverse dynamics model unit 12. In this case as well, the hand load estimation device 1 can obtain a highly accurate estimation result of the hand load parameter. In the hand load estimation device 1 shown in FIG. 2 as well, the detected torque value may be used as the value of the actual torque τ _real .

Although the neural network used by the learning unit 13a is an RNN, this is not limiting. The learning unit 13a may use a feedforward type, i.e., forward propagation type, neural network. That is, the learning unit 13a may perform supervised learning of the hand load parameters using a neural network in which the output is determined only by the input of time series data of the position θ, velocity θ′, acceleration θ″, and difference τ _real -τ _mdl of each joint, instead of a neural network in which the output of the intermediate layer 22 at time t is recursively input to the intermediate layer 22 at the next time. In this case, the learning unit 13a inputs data to the neural network after a predetermined batch size of data is accumulated, rather than sequentially inputting data to the neural network every time data is acquired at time t.

In the first embodiment, the learning unit 13, 13a learns the hand load parameter from which the influence of the modeling error and the influence of the out-of-model characteristic are removed. That is, the learning unit 13, 13a learns the hand load parameter and the torque error corresponding to the hand load parameter together by using one neural network. The learning unit 13, 13a may separate the neural network for learning the hand load parameter from the neural network for learning the torque error. In this case, the neural network for learning the hand load parameter is constructed so that the hand load parameter output by inputting the position θ, the velocity θ', the acceleration θ'' and the difference τ _real -τ _mdl coincides with the correct label. The neural network for learning the torque error is constructed so that the torque error output by inputting the position θ, the velocity θ', the acceleration θ'' and the hand load parameter coincides with the correct label, the difference τ _real -τ _mdl . The hand load estimation device 1 corrects the inference result of the hand load parameter based on the inference result of the torque error. In this case, too, the hand load estimation device 1 can estimate the hand load parameter from which the influence of the modeling error and the influence of the characteristics outside the model are removed, and can obtain a highly accurate estimation result of the hand load parameter.

Embodiment 2.
In the second embodiment, a case will be described in which learning is performed using data when no load is applied to the hand using GPR (Gaussian Process Regression). Hereinafter, no load is applied to the hand, which is referred to as "no load." Fig. 6 is a diagram showing the configuration of a hand load estimation device 2 according to the second embodiment. In the second embodiment, the same components as those in the first embodiment are given the same reference numerals, and the configuration different from the first embodiment will be mainly described.

The hand load estimation device 2 has an input unit 11, inverse dynamics model units 12, 12b, a learning unit 13b, a learning result storage unit 14b, and an inference unit 15b. When the learning unit 13b performs learning, the inverse dynamics model unit 12 calculates a nominal torque τ mdl from the position θ, velocity θ', and acceleration θ'' at no load, based on an inverse dynamics model in which nominal parameters are set. The position θ, velocity θ', and acceleration θ'' at no load, and the difference τ _real -τ _mdl at no _load are input to the learning unit 13b as learning data. The learning unit 13b learns hand load parameters from which modeling errors at no load and the influence of characteristics outside the model have been removed, using the learning data. The learning unit 13b stores the learning results in the learning result storage unit 14b.

During inference by the inference unit 15b, the inverse dynamics model unit 12b calculates an estimated torque τ est assuming no load from the position θ, velocity θ', and acceleration θ'', which are inference data, based on the corrected inverse dynamics model. The position θ, velocity θ', and acceleration θ'' of each joint and _the difference τ _real -τ _est are input to the inference unit 15b as inference data. The inference unit 15b also reads out the learning results by the learning unit 13b from the learning result storage unit 14b. The inference unit 15b estimates hand load parameters based on the learning results by the learning unit 13b.

Figure 7 is a diagram for explaining GPR used for learning in the second embodiment. GPR is a method used to estimate a nonlinear function y = f(x). In GPR, it is assumed that the observed value y follows a multidimensional Gaussian distribution N(u, S), and the internal parameter z of the kernel function S is adjusted so that the likelihood p(f|z) in Bayesian estimation is maximized. The Gaussian distribution N(u, S) is determined by a mean function u(x) corresponding to a certain average value and a kernel function S(x, x') corresponding to a certain variance value. The estimated function f~N(u, S) is output as a distribution of functions. In a region where the kernel function S is large, the uncertainty of the output y for the input x becomes large. The following formula (8) represents the estimated function f~N(u, S).

The learning unit 13b equipped with a GPR receives as input the position θ, velocity θ', and acceleration θ" at no load, which are learning data, and learns the distribution of a function that outputs the difference τ _real -τ _mdl . The learning unit 13b stores the obtained distribution of the function in the learning result storage unit 14b. The learning data may be only a portion of the data among the position θ, velocity θ', and acceleration θ". Of the position θ, velocity θ', and acceleration θ", for example, only the data of velocity θ' may be input to the learning unit 13b as learning data. In this way, the learning unit 13b may reduce the dimension of the input data by limiting the input data to only a portion of the data among the position θ, velocity θ', and acceleration θ".

The inference unit 15b may obtain the hand load parameter estimation result using a histogram filter, which is a type of Bayes filter. In the histogram filter, the existence range of the parameter to be estimated is divided into specific intervals, and candidate points are set for each interval, and the existence probability of each candidate point is updated based on the time series data.

8 is a diagram for explaining a histogram filter used for inference in the second embodiment. The inverse dynamics model unit 12b reads out the distribution of the function, which is the learning result of the GPR, from the learning result storage unit 14b, and corrects the inverse dynamics model based on the distribution of the function. The inverse dynamics model unit 12b inputs the position θ, velocity θ', and acceleration θ'', which are inference data, to the inverse dynamics model corrected based on the learning result of the GPR, and outputs an estimated torque τ _est in the case where no load is assumed.

The difference τ _real - τ _est between the real torque τ _real and the estimated torque τ _est of the inference data when the robot is operated with a load applied to the hand can be considered as the offset torque generated by the load. Hereinafter, the offset torque will be referred to as the offset torque Δτ. The inference unit 15b prepares candidate points for the parameters to be estimated, and outputs the candidate point with the highest probability of obtaining the actual offset torque Δτ data for the position θ, velocity θ', and acceleration θ'' as the hand load parameter estimation result.

In Fig. 8, the parameter to be estimated is one type, for example, only the mass of the hand. Fig. 8 shows an example in which the probability of obtaining the input offset torque Δτ data is highest at candidate point m _i , which is one of candidate points m ₁ , m ₂ , ..., m _j obtained by equally dividing the range of possible mass values into j parts. j is an arbitrary integer, and i is an integer satisfying 1≦i≦j. When there are K types of parameters to be estimated, ^jK candidate points can be prepared by dividing the possible range of each parameter into j parts. K is an integer of 2 or more.

The step size of the candidate points may differ depending on the type of parameter to be estimated. When updating the probability of existence of each candidate point by inputting data on the position θ, velocity θ', and offset torque Δτ for acceleration θ'' at time t, the inference unit 15b may change the possible range of each parameter to be estimated or the step size of each parameter.

If the existence probability of a candidate point m _i of the histogram filter is p _i , the expected value of the amount of information of the entire histogram filter is expressed by the following formula (9).

H in equation (9) is called entropy. Entropy H has a positive value. The closer entropy H is to zero, the more the histogram filter has converged to a candidate point. To avoid adverse effects due to outliers in the inference data, the inference unit 15b may calculate entropy H each time data at time t is input, and may consider the data to be valid only if entropy H is lower than the previously calculated value of entropy H, and update the existence probability of each candidate point.

According to the second embodiment, the hand load estimation device 2 learns the hand load parameters based on the learning data at no load, and infers the hand load parameters based on the actual torque and the estimated torque at no load. This allows the hand load estimation device 2 to obtain highly accurate estimation results of the hand load parameters.

In the second embodiment, the hand load estimation device 2 performs learning using data obtained when the robot is operated without a load, but the present invention is not limited to this. The hand load estimation device 2 may perform learning using data obtained when the robot is operated with a load, the hand load parameters of which are known, attached to the hand, such as the mass or the center of gravity position. That is, instead of learning the difference τ _real -τ mdl between the actual torque τ _real and the nominal torque τ _mdl calculated based on the parameters when no load is applied, the hand load estimation device 2 learns the difference τ _real -τ _mdl between the actual torque τ _real and the nominal torque τ _mdl calculated based on the known hand load parameters, using the GPR. The hand load estimation device 2 can estimate the hand _load parameters using a histogram filter from the difference τ _real -τ _est between the actual torque τ _real of the inference data and the estimated torque τ _est when it is assumed that the same load as that used during learning is attached to the hand.

In the learning in the second embodiment, a learning method of an approximation function instead of GPR may be applied to learn a nonlinear unknown function. For example, a method such as Neural Processes or GP Net may be applied to the learning. Neural Processes is known as a method of learning a neural network that generates a model of a stochastic process from data. GP Net is known as a method of approximating a predicted distribution by data using a Bayesian neural network, assuming that the prior distribution of an unknown function follows a Gaussian process. In these methods, the mean value and variance value of a function can be output, as in the case of GPR. In addition, in the inference in the second embodiment, a method such as Particle Swarm Optimization (PSO), Genetic Algorithm (GA), or Simulated Annealing (SA), which are known as methods for solving combinatorial optimization problems, may be applied to update the candidate points of the hand load parameters to match the data of the offset torque Δτ.

In the second embodiment, in order to learn a nonlinear unknown function, learning by GPR may be combined with other learning. For example, data input in learning by GPR is used as input to a neural network, and learning is performed using the average value and variance value of the function output by GPR as teacher signals. In this case, the hand load estimation device 2 can output the average value and variance value at the time of inference using the trained neural network.

Embodiment 3.
In the third embodiment, a case where joint temperature data is used for learning will be described. Fig. 9 is a diagram showing the configuration of a hand load estimating device 3 according to the third embodiment. In the third embodiment, the same components as those in the first or second embodiment are given the same reference numerals, and the configuration different from the first or second embodiment will be mainly described.

The hand load estimation device 3 has an input unit 11, an inverse dynamics model unit 12, a learning unit 13c, a learning result storage unit 14c, and an inference unit 15c. A position FB value, which is position data, a current FB value, which is current data, and a temperature feedback (FB) value, which is temperature data, are input to the input unit 11. The temperature FB value is the temperature of each joint when the robot is operating.

The learning unit 13c receives as learning data the position θ, velocity θ', and acceleration θ'' of each joint based on the input data during learning, the difference τ _real -τ _mdl of each joint based on the input data during learning, and temperature data of each joint. Using the learning data, the learning unit 13c learns hand load parameters from which the effects of the temperature dependency of friction characteristics in the joints have been removed. The learning unit 13c stores the learning results in the learning result storage unit 14c.

The inference unit 15c receives, as inference data, the position θ, velocity θ', and acceleration θ'' of each joint based on the input data at the time of inference, the difference τ _real -τ _mdl of each joint based on the input data at the time of inference, and temperature data of each joint. Furthermore, the inference unit 15c reads out the learning results by the learning unit 13c from the learning result storage unit 14c. The inference unit 15c estimates hand load parameters based on the learning results by the learning unit 13c.

The temperature dependency of frictional characteristics is one of the factors that cause modeling errors or characteristics outside the model. According to the third embodiment, the hand load estimation device 3 learns hand load parameters from which the influence of the temperature dependency of frictional characteristics has been removed, and infers the hand load parameters based on the learning results. The hand load estimation device 3 is capable of inferring hand load parameters from which the influence of the temperature dependency of frictional characteristics has been reduced, and is therefore able to obtain highly accurate estimation results of the hand load parameters. Note that the same learning and inference as in the third embodiment may be applied to the hand load estimation devices 1 and 2 of the first and second embodiments.

Embodiment 4.
In the fourth embodiment, a case will be described in which motion data, which is data on the position θ, velocity θ', and acceleration θ'' of each joint, is converted into variables. Fig. 10 is a diagram showing the configuration of a hand end load estimating device 4 according to the fourth embodiment. In the fourth embodiment, the same components as those in the first to third embodiments are given the same reference numerals, and the configuration different from the first to third embodiments will be mainly described.

The hand load estimation device 4 has an input unit 11, an inverse dynamics model unit 12, a learning unit 13d, a learning result storage unit 14d, an inference unit 15d, and a variable conversion unit 16. The position θ, velocity θ', and acceleration θ'' calculated from the position data are input to the variable conversion unit 16. The variable conversion unit 16 converts the position θ, velocity θ', and acceleration θ'' into variables for setting a torque equation linearized with respect to the hand load parameters. The variable conversion unit 16 outputs the converted variables.

During learning by the learning unit 13d, the variable conversion unit 16 converts the data of position θ, velocity θ', and acceleration θ'' obtained from the input data during learning into variables, and outputs the variables to the learning unit 13d. During inference by the inference unit 15d, the variable conversion unit 16 converts the data of position θ, velocity θ', and acceleration θ'' obtained from the input data during inference into variables, and outputs the variables to the inference unit 15d. Note that in Figure 10, the variable conversion unit 16 is shown divided into two for the convenience of showing the data flow during learning and the data flow during inference separately.

Variables converted from data on the position θ, velocity θ', and acceleration θ'' of each joint, and the difference τ _real -τ _mdl of each joint are input to the learning unit 13d as learning data. The learning unit 13d uses the learning data to learn hand load parameters from which the effects of modeling errors in the inverse dynamics model and the effects of out-of-model characteristics have been removed. The learning unit 13d stores the learning results in the learning result storage unit 14d.

The inference unit 15d receives, as inference data, variables converted from the data on the position θ, velocity θ', and acceleration θ'' of each joint, and the difference τ _real -τ _mdl of each joint. The inference unit 15d reads out the learning results of the learning unit 13d from the learning result storage unit 14d. The inference unit 15d estimates hand load parameters based on the learning results of the learning unit 13d.

The equation of motion for a robot is a nonlinear equation with respect to position θ, velocity θ', and acceleration θ''. However, if a parameter vector M is an arrangement of parameters such as the mass and center of gravity of the load in the form (mass) x (center of gravity), the equation of motion can be linearized as shown in the following equation (10). K(θ, θ', θ'') is called the regressor matrix. K(θ, θ', θ'') is a variable converted from the position θ, velocity θ', and acceleration θ'' of each joint.

The RNN and GPR described above can handle processing by nonlinear regression. By converting the data of the position θ, velocity θ', and acceleration θ'' into variables, the hand load estimation device 4 can handle processing handled by nonlinear regression as processing in a linear domain. This enables the hand load estimation device 4 to learn and infer using simple processing. The hand load estimation device 4 can obtain highly accurate estimation results of hand load parameters using simple processing. The same variable conversion as in embodiment 4 may be applied to the hand load estimation devices 1, 2, and 3 of embodiments 1 to 3. Note that in embodiment 7 described later, an example of applying a regressor matrix to inference by combining sequential identification of hand load parameters and probability distribution will be described.

Embodiment 5.
In the fifth embodiment, a case will be described in which information on friction characteristics related to the position θ, velocity θ', or acceleration θ'' of each joint is used for learning. FIG. 11 is a diagram showing the configuration of a hand end load estimating device 5 according to the fifth embodiment. In the fifth embodiment, the same components as those in the first to fourth embodiments are given the same reference numerals, and the configuration different from the first to fourth embodiments will be mainly described.

The hand load estimation device 5 has an input unit 11, an inverse dynamics model unit 12, a learning unit 13e, a learning result storage unit 14e, and an inference unit 15e. The position θ, velocity θ', and acceleration θ'' of each joint based on the input data at the time of learning, and the difference τ _real -τ _mdl of each joint based on the input data at the time of learning are input to the learning unit 13e as learning data.

In addition, information on friction characteristics is input to the learning unit 13e. The information on friction characteristics input to the learning unit 13e is information identified in advance separately from the learning data, and is information related to at least one of the position θ, velocity θ', and acceleration θ'' of each joint. Note that the friction characteristics being related to at least one of the position θ, velocity θ', and acceleration θ'' refers to the friction characteristics changing in response to a change in at least one of the position θ, velocity θ', and acceleration θ''.

The learning unit 13e learns hand load parameters from which the influence of friction characteristics related to the position θ, velocity θ', or acceleration θ'' of each joint has been removed by learning using the learning data and information on friction characteristics. The learning unit 13e stores the learning results in the learning result storage unit 14e. The inference unit 15e reads out the learning results by the learning unit 13e from the learning result storage unit 14e. The inference unit 15e estimates the hand load parameters based on the learning results by the learning unit 13e.

Friction characteristics related to position θ, velocity θ', or acceleration θ'' are cited as one of the causes of modeling error or characteristics outside the model. According to embodiment 5, the hand load estimation device 5 learns hand load parameters from which the influence of friction characteristics related to position θ, velocity θ', or acceleration θ'' of each joint has been removed, and infers the hand load parameters based on the learning results. The hand load estimation device 5 is capable of inferring hand load parameters from which the influence of friction characteristics related to position θ, velocity θ', or acceleration θ'' has been reduced, thereby making it possible to obtain highly accurate estimation results of the hand load parameters.

Information on friction characteristics related to the temperature of each joint may be input to the learning unit 13e. The information on friction characteristics related to the temperature of each joint is information that is identified in advance separately from the learning data. In this case, the learning unit 13e learns hand load parameters from which the influence of the friction characteristics related to the temperature of each joint has been removed. The friction characteristics being related to temperature refers to the friction characteristics changing in response to a change in temperature. In this case as well, the hand load estimation device 5 can obtain highly accurate estimation results of the hand load parameters. In addition, learning similar to that of the fifth embodiment may be applied to the hand load estimation devices 1, 2, 3, and 4 of the first to fourth embodiments.

Embodiment 6.
In the sixth embodiment, a case where an index indicating the reliability of a hand load parameter is calculated and output will be described. Fig. 12 is a diagram showing the configuration of a hand load estimating device 6 according to the sixth embodiment. In the sixth embodiment, the same components as those in the first to fifth embodiments are given the same reference numerals, and the configuration different from the first to fifth embodiments will be mainly described.

The hand load estimation device 6 has an input unit 11, an inverse dynamics model unit 12, a learning unit 13f, a learning result storage unit 14f, and an inference unit 15f. The learning unit 13f uses learning data to learn hand load parameters from which the effects of temperature dependency of friction characteristics in joints have been removed. The learning unit 13f stores the learning results in the learning result storage unit 14f. The inference unit 15f reads out the learning results by the learning unit 13f from the learning result storage unit 14f. The inference unit 15f infers hand load parameters based on the learning results by the learning unit 13f. Furthermore, the learning unit 13f calculates a reliability index and stores the calculation result in the learning result storage unit 14f. The inference unit 15f reads out the reliability index from the learning result storage unit 14f, and outputs the reliability index together with the inference result of the hand load parameter.

When performing inference using a neural network, the reliability index can be obtained by using a measure that represents the difference between the probability distribution at the time of learning and the probability distribution at the time of inference. As the measure, there are JS divergence (Jensen-Shannon divergence), KL divergence (Kullback-Leibler divergence), etc., and any of them may be used. The learning unit 13f sets the mean value and variance for each parameter as the output of the neural network. Here, the measure that represents the difference is the difference between the probability distribution of the correct label and the probability distribution of the inference value expressed by the mean value and variance obtained as the output of the neural network, and is a value obtained by calculating using KL divergence or the like. The correct label is a teacher signal for the mean value. The learning unit 13f performs learning so as to bring the mean value and variance of each parameter closer to the probability distribution of the correct label by constructing a neural network. The hand load estimation device 6 treats the mean value of each parameter as an estimated value and the variance as an index of reliability. The smaller the variance, the higher the reliability of the estimated value is evaluated.

When making inference using a neural network, the output result from Dropout can also be used as a reliability index. Dropout is known as a technique for learning by randomly setting the use and non-use of units at a certain ratio for each unit of a neural network. In this case, several neural networks are prepared in which units are randomly used or not used at the same ratio during inference, and estimation is performed with each of them, ensuring variation in the estimation. The average value of multiple estimation results is treated as the estimate, and the variance is treated as a reliability index, and the smaller the variance, the higher the reliability of the estimate is evaluated.

When making inference using a histogram filter, the magnitude of the certainty of the histogram filter can be used as a reliability index. In this case, the greater the probability of the candidate point that shows the greatest existence probability when normalized so that the sum of the existence probabilities at all candidate points is 1, the higher the reliability of the estimated value is evaluated to be. It can also be evaluated that the smaller the value of the entropy H of the histogram filter calculated by the above formula (9), the higher the reliability of the estimated value.

When learning using GPR, the variance of the distribution of the function obtained as a result of learning can be used as a reliability index. In this case, the smaller the variance of the distribution of the torque error function shown by GPR for the region of position θ, velocity θ', and acceleration θ'' at the time of inference, the higher the reliability of the estimated value is evaluated to be.

According to the sixth embodiment, the hand load estimation device 6 calculates an index indicating the reliability of the hand load parameter, and outputs the reliability index together with the inference result of the hand load parameter, thereby making it possible to evaluate the reliability of the estimated hand load parameter.

Embodiment 7.
In the seventh embodiment, an example will be described in which the inference unit outputs a reliability index together with the estimated value of the hand load parameter. Fig. 13 is a diagram showing the configuration of a hand load estimation device 7 according to the seventh embodiment. In the seventh embodiment, the same components as those in the first to sixth embodiments are given the same reference numerals, and the configuration different from the first to sixth embodiments will be mainly described.

The hand load estimation device 7 has an input unit 11, inverse dynamics model units 12, 12g, a learning unit 13g, a learning result storage unit 14g, an inference unit 15g, and a variable conversion unit 16. When the learning unit 13g is learning, the variable conversion unit 16 converts the position θ, velocity θ', and acceleration θ'', which are movement data obtained from the input data during learning, into variables, and outputs the variables to the learning unit 13g. When the inference unit 15g is making an inference, the variable conversion unit 16 converts the position θ, velocity θ', and acceleration θ'', which are obtained from the input data during inference, into variables, and outputs the variables to the inference unit 15g.

The learning unit 13g receives K(θ, θ', θ'') and the difference τ _real -τ _mdl as learning data. The regressor matrix K(θ, θ', θ'') is a variable converted from the position θ, velocity θ', and acceleration θ'', as described in the fourth embodiment. Furthermore, information on nominal parameters is input to the learning unit 13g. The learning unit 13g equipped with a GPR receives the position θ, velocity θ', and acceleration θ'' at no load, which are learning data, as input, and learns the distribution of a function that outputs the difference τ _real -τ _mdl . The learning unit 13g stores the distribution of the function, which is the learning result, in the learning result storage unit 14g.

The inverse dynamics model unit 12g reads out the distribution of the function, which is the learning result of the GPR, from the learning result storage unit 14g, and corrects the inverse dynamics model based on the distribution of the function. The inverse dynamics model unit 12g outputs an estimated torque τ est in the case where no load is assumed, by inputting the position θ, velocity θ', and acceleration θ'' based on the position data at the time of inference to the inverse dynamics _model corrected by the learning result of the GPR.

The inference unit 15g receives K(θ, θ', θ'') and the offset torque Δτ as inference data. The offset torque Δτ is the difference τ _real - τ _est between the actual torque τ _real and the estimated torque τ _est . The inference unit 15g reads out the learning result by the learning unit 13g from the learning result storage unit 14g. The inference unit 15g estimates the hand load parameters based on the learning result by the learning unit 13g.

K(θ, θ', θ'') and the parameter vector M in the above formula (10) can be expanded to include a term related to friction torque. The learning unit 13g receives as input Y, which is a vector in which the real torque τ _real at each time included in the no-load data for learning is arranged, K, which is a regret matrix calculated from the position θ, velocity θ', and acceleration θ'' at each time by the variable conversion unit 16, and the nominal parameters included in the inverse dynamics model, and calculates an optimal solution of the parameter vector M such that the nominal torque calculated by the above formula (10) at no-load matches the actual torque. The optimal solution of the parameter vector M is expressed by the following formula (11). In formula (11), M with "^" above it represents the optimal solution of the parameter vector M. M with "￣" above it represents the nominal parameters. T represents a transposed matrix. W represents a weighting matrix.

The following equation (12) holds for the offset torque Δτ and the hand load parameter vector m to be estimated. This allows the inference unit 15g to estimate the hand load parameter vector m using any linear regression method.

When Bayesian linear regression is used as the linear regression method, the inference unit 15g can output an indicator of the reliability of the estimation result together with the estimation result of the hand load parameters.

FIG. 14 is a conceptual diagram for explaining the update of the probability distribution by Bayesian linear regression in the seventh embodiment. The prior probability distribution N(m|u ₀ , S ₀ ) is updated to the posterior probability distribution N(m|u, S) by the observation data Δτ, K(θ, θ′, θ″). Such updating is repeated. In the Bayesian linear regression, the hand load parameter vector m when the data at time t is input is assumed to follow the probability distribution N(m|u, S), and the mean u and variance S are updated according to the following formulas (13) and (14), respectively. u ₀ represents the prior probability distribution of the mean. S ₀ represents the prior probability distribution of the variance. β is a parameter representing the inverse of the covariance of the torque error, and is set according to the noise distribution included in the torque data.

The inference unit 15g outputs the mean u of the probability distribution as an estimate of the hand load parameter. The inference unit 15g also outputs the variance S of the probability distribution as an index of reliability. The smaller the variance S, the higher the reliability of the estimate is evaluated.

According to the seventh embodiment, the hand load estimation device 7 outputs an index of reliability together with the inference result of the hand load parameter, thereby making it possible to evaluate the reliability of the estimated hand load parameter.

Embodiment 8.
FIG. 15 is a diagram showing a configuration of a robot system 8 according to an eighth embodiment. The robot system 8 includes a robot control system 9 and a robot 40. The robot control system 9 includes a hand load estimation device 1 and a robot control device 30 that controls the robot 40. In the eighth embodiment, the robot control device 30 adjusts the acceleration/deceleration of the robot 40 based on the hand load parameters output from the hand load estimation device 1 during the operation of the robot 40. In the eighth embodiment, the same components as those in the first to seventh embodiments are denoted by the same reference numerals, and the configuration different from the first to seventh embodiments will be mainly described. The robot system 8 may include the hand load estimation device according to any one of the first to seventh embodiments.

The robot controller 30 has a command generator 31. The robot controller 30 receives the hand load parameters output from the hand load estimation device 1. The command generator 31 generates a command to be sent to the robot 40 based on the hand load parameters output from the hand load estimation device 1. The robot controller 30 sends the generated command to the robot 40.

The hand load estimation device 1 is communicatively connected to the robot control device 30. The hand load estimation device 1 is not limited to being an external device to the robot control device 30. The hand load estimation device 1 may be built into the robot control device 30. The learning and inference in any one of the first to seventh embodiments may be performed inside the robot control device 30.

Alternatively, of the learning unit 13 and the inference unit 15 of the hand load estimation device 1, the learning unit 13 may be realized by a device external to the robot control device 30, and the inference unit 15 may be built into the hand load estimation device 1 inside the robot control device 30. In this case, a learning device having an input unit 11, an inverse dynamics model unit 12, a learning unit 13, and a learning result storage unit 14 is provided outside the robot control device 30. Also, an input unit 11 and an inverse dynamics model unit 12 similar to the learning device, and an inference unit 15 that performs inference based on a learned model read from the learning result storage unit 14 of the learning device are provided in the hand load estimation device 1 inside the robot control device 30. In this way, the learning in any one of the embodiments 1 to 7 may be performed outside the robot control device 30, and the inference in any one of the embodiments 1 to 7 may be performed inside the robot control device 30.

Next, the change in acceleration/deceleration of the robot 40 by the robot controller 30 will be described. The robot controller 30 has a function of calculating the shortest acceleration time and deceleration time that do not exceed the allowable value of the torque that operates each joint. The robot controller 30 determines the acceleration time and deceleration time by using the hand load parameters to calculate the equation of motion for each of the acceleration start point, acceleration end point, deceleration start point, and deceleration end point of the robot 40. The robot controller 30 changes the acceleration/deceleration speed based on the acceleration time and deceleration time.

The robot control device 30 calculates the deceleration start point and deceleration end point using the hand load parameters estimated from the data up to the acceleration/deceleration end point. In this case, it is possible to shorten the deceleration time compared to when the deceleration time is calculated based on the hand load parameters that are set to a conservative fixed value so that the actual torque does not exceed the allowable value. By being able to shorten the deceleration time without exceeding the allowable torque value, the robot control device 30 can achieve an increase in the speed of the robot 40 without excessive torque.

When an index indicating the reliability of the estimation result is obtained along with the hand load parameter estimation result, the robot controller 30 may further adjust the acceleration/deceleration rate according to the reliability index. For example, the robot controller 30 may reflect the reliability in the value of the hand load parameter used in the calculation of the optimal acceleration/deceleration. In this case, the robot controller 30 may correct the hand load parameter in a conservative direction according to the reliability, within a range in which the deceleration time is not longer than when the hand load parameter is a fixed value.

The robot controller 30 may reflect the reliability in the deceleration time calculated using the estimated hand load parameter value as is. In this case, the robot controller 30 uses the deceleration time calculated using the estimated hand parameter value as is multiplied by a correction coefficient a (where a>1) according to the reliability as the deceleration time of the robot 40.

The robot control device 30 may use the estimated hand load parameters when calculating the acceleration and deceleration times of the next operation that is continued under load after one operation for obtaining data used to estimate the hand load parameters is completed.

According to the eighth embodiment, the robot system 8 can increase the speed of the robot 40 by changing the acceleration/deceleration of the robot based on the hand load parameters.

Embodiment 9.
In the ninth embodiment, an example of adjusting parameters for detecting a collision of the robot 40 in the robot system 8 will be described with reference to Fig. 15. The robot control device 30 adjusts a threshold value, which is a parameter for detecting a collision of the robot 40, based on the hand load parameter while the robot 40 is in operation. The threshold value is a threshold value for determining whether or not the robot 40 has collided with an obstacle.

The robot controller 30 has a function of detecting a collision by comparing the actual torque and estimated torque of each joint. If the actual torque is greater than the estimated torque and the difference between the actual torque and the estimated torque is greater than a preset threshold, the robot controller 30 determines that the robot 40 has collided with an obstacle. A margin is set in advance for the threshold to prevent erroneous detection due to an error between the actual torque and the estimated torque when the robot 40 is operating normally. In addition, the robot controller 30 can increase the collision detection sensitivity by adjusting the threshold based on the hand load parameters compared to when the threshold is fixed.

When an index indicating the reliability of the estimation result is obtained together with the hand load parameter estimation result, the robot control device 30 may further adjust the threshold value according to the reliability index. For example, the robot control device 30 can reflect the reliability in the value of the hand load parameter used to calculate the estimated torque. In this case, the robot control device 30 can adjust the threshold value in a conservative direction, so to speak, by adjusting the threshold value based on the hand load parameter corrected according to the reliability.

The robot controller 30 may reflect the reliability of the threshold calculated using the estimated hand load parameter value as is. In this case, the robot controller 30 calculates the estimated torque using the estimated hand parameter value as is, and adjusts the threshold by multiplying the threshold calculated based on the estimated torque by a correction coefficient b (where b>1) according to the reliability.

As in embodiment 8, the hand load estimation device 1 is not limited to being a device external to the robot control device 30. The hand load estimation device 1 may be built into the robot control device 30. The learning and inference in any one of embodiments 1 to 7 may be performed inside the robot control device 30. Alternatively, the learning in any one of embodiments 1 to 7 may be performed outside the robot control device 30, and the inference in any one of embodiments 1 to 7 may be performed inside the robot control device 30.

According to the ninth embodiment, the robot system 8 can reduce collisions of the robot 40 with obstacles and improve the safety of the robot 40 by adjusting the parameters for collision detection based on the hand load parameters.

Next, a hardware configuration for realizing the hand load estimation device 1, 2, 3, 4, 5, 6, 7 according to the first to seventh embodiments will be described. FIG. 16 is a diagram showing a first example of the hardware configuration for realizing the hand load estimation device 1, 2, 3, 4, 5, 6, 7 according to the first to seventh embodiments. The first example of the configuration is a configuration example in which the inverse dynamics model unit 12, 12b, 12g, the learning unit 13, 13a, 13b, 13c, 13d, 13e, 13f, 13g, the inference unit 15, 15a, 15b, 15c, 15d, 15e, 15f, 15g, and the variable conversion unit 16, which are the main parts of the hand load estimation device 1, 2, 3, 4, 5, 6, 7, are realized by a processing circuit 51 having a processor 53 and a memory 54.

The input unit 52 is a circuit that receives an input signal for the hand load estimation device 1, 2, 3, 4, 5, 6, 7 from the outside. The input unit 11 shown in FIG. 1 etc. is realized by the input unit 52. The output unit 55 is a circuit that outputs a signal generated by the hand load estimation device 1, 2, 3, 4, 5, 6, 7 to the outside. The output unit 55 outputs the hand load parameters that are the estimation results and the reliability index in the sixth and seventh embodiments.

The processor 53 is a CPU (Central Processing Unit). The processor 53 may be an arithmetic unit, a microprocessor, a microcomputer, or a DSP (Digital Signal Processor). The memory 54 is, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), or an EEPROM (registered trademark) (Electrically Erasable Programmable Read Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disk).

The processor 53 executes the hand load estimation program. The hand load estimation program is a program that describes the processing for each part constituting the main part of the hand load estimation device 1, 2, 3, 4, 5, 6, 7. The hand load estimation program is stored in advance in the memory 54. The processor 53 reads out and executes the hand load estimation program stored in the memory 54, thereby operating as each part constituting the main part of the hand load estimation device 1, 2, 3, 4, 5, 6, 7. The memory 54 also stores various information. The learning result storage parts 14, 14a, 14b, 14c, 14d, 14e, 14f, 14g are realized by the memory 54. The memory 54 is also used as a temporary memory when the processor 53 executes various processes.

The hand load estimation program is assumed to be stored in advance in memory 54, but is not limited to this. The hand load estimation program may be provided to a user of the hand load estimation device 1, 2, 3, 4, 5, 6, 7 in a state written on a storage medium that can be read by a computer system, and may be installed in memory 54 by the user. The storage medium may be a portable storage medium such as a flexible disk, or a flash memory such as a semiconductor memory. The hand load estimation program may be installed in memory 54 from another computer or server device via a communication network.

The main parts of the hand load estimation devices 1, 2, 3, 4, 5, 6, and 7 according to the first to seventh embodiments may be realized by dedicated hardware. FIG. 17 is a diagram showing a second configuration example of the hardware that realizes the hand load estimation devices 1, 2, 3, 4, 5, 6, and 7 according to the first to seventh embodiments. The second configuration example is a configuration example in which the function of the processing circuit 51 shown in FIG. 16 is realized by the processing circuit 56, which is dedicated hardware.

The processing circuit 56 is, for example, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a circuit that combines these. In the example shown in FIG. 17, the main parts of the hand load estimation devices 1, 2, 3, 4, 5, 6, and 7 are realized by a single processing circuit 56, but this is not limited to this. The hardware may include multiple processing circuits 56, and the main parts of the hand load estimation devices 1, 2, 3, 4, 5, 6, and 7 may be realized by different processing circuits 56. The storage device 57 is a hard disk drive (HDD) or a solid state drive (SSD) that stores various information. The learning result storage units 14, 14a, 14b, 14c, 14d, 14e, 14f, and 14g are realized by the storage device 57.

In addition, some of the main parts of the hand load estimation device 1, 2, 3, 4, 5, 6, and 7 may be realized by the processor 53 and memory 54 shown in FIG. 16, and the rest may be realized by dedicated hardware similar to the processing circuit 56.

The configurations shown in the above embodiments are examples of the contents of this disclosure. The configurations of each embodiment can be combined with other known technologies. The configurations of each embodiment can be combined as appropriate. A portion of the configuration of each embodiment can be omitted or modified without departing from the gist of this disclosure.

1, 2, 3, 4, 5, 6, 7 Hand load estimation device, 8 Robot system, 9 Robot control system, 11, 52 Input section, 12, 12b, 12g Inverse dynamics model section, 13, 13a, 13b, 13c, 13d, 13e, 13f, 13g Learning section, 14, 14a, 14b, 14c, 14d, 14e, 14f, 14g Learning result storage section, 15, 15a, 15b, 15c, 15d, 15e, 15f, 15g Inference section, 16 Variable conversion section, 21 Input layer, 22 Intermediate layer, 23 Output layer, 24, 25, 26 Neuron, 27 Torque sensor, 30 Robot control device, 31 Command generation section, 40 Robot, 51, 56 Processing circuit, 53 Processor, 54 Memory, 55 Output section, 57 storage device.

Claims (20)

A hand load estimation device that estimates a hand load parameter which is a parameter regarding a load applied to a hand of a robot, comprising:
an inverse dynamics model unit that outputs a nominal torque value, which is a torque calculated based on an inverse dynamics model, by inputting motion data, which is at least one of data on a position, a velocity, and an acceleration of a joint of the robot, into the inverse dynamics model;
a learning unit that learns the hand load parameters from which an influence of a modeling error of the inverse dynamics model and an influence of a characteristic of the robot that is not included in the inverse dynamics model are removed, based on learning data including the motion data and a difference between the actual torque and the nominal torque of the joint;
an inference unit to which inference data including the motion data and the difference between the actual torque and the nominal torque is input, and which infers the hand load parameter based on a learning result by the learning unit;
A hand load estimation device comprising: Position data representing positions of the joints is input to the hand load estimation device,
2. The hand load estimating device according to claim 1, wherein the movement data obtained from the position data is input to the inverse dynamics model section, the learning section and the inference section. The hand load estimation device according to claim 2, characterized in that the position data is a command value for controlling the joint or a feedback value of the position of the joint. The hand load estimation device according to any one of claims 1 to 3, characterized in that the actual torque value for calculating the difference is a value obtained by multiplying the current data of the joint by a torque constant. The hand load estimation device according to any one of claims 1 to 3, characterized in that the actual torque value for calculating the difference is a torque value detected by a torque sensor provided on the robot. A hand load estimation device according to any one of claims 1 to 5, characterized in that the movement data when a load is applied to the hand, the difference calculated based on the actual torque and the nominal torque when a load is applied to the hand, and the correct answer label of the hand load parameter are input to the learning unit. The motion data when no load is applied to the hand and the difference calculated based on the actual torque and the nominal torque when no load is applied to the hand are input to the learning unit,
6. The hand load estimation device according to claim 1, wherein the inference unit infers the hand load parameter based on a difference between the actual torque and an estimated torque, which is the nominal torque when no load is added to the hand. The motion data when a load having a known hand load parameter is applied to the hand, and the difference calculated based on the actual torque and the nominal torque when a load having a known hand load parameter is applied to the hand are input to the learning unit,
The hand load estimation device according to any one of claims 1 to 5, characterized in that the inference unit infers the hand load parameter based on a difference between the actual torque and an estimated torque, which is the nominal torque when a load, the hand load parameter of which is known, is applied to the hand. The hand load estimation device according to any one of claims 1 to 8, characterized in that the learning unit receives input of information on friction characteristics related to at least one of the position, velocity, and acceleration of the joint. The hand load estimation device according to any one of claims 1 to 5, characterized in that the learning unit learns the hand load parameters that do not include errors due to temperature dependency of friction characteristics in the joint by inputting temperature data indicating the temperature of the joint to the learning unit. The hand load estimation device according to claim 10, characterized in that the learning unit receives input of information on friction characteristics related to at least one of the position, velocity, acceleration, and temperature of the joint. A variable conversion unit converts the motion data into variables for setting a linearized torque equation with respect to the hand load parameters,
12. The hand load estimating device according to claim 1, wherein the variables are input to the learning section and the inference section. The learning unit calculates an index indicating a reliability of the hand load parameter,
13. The hand load estimating device according to claim 1, wherein the inference unit outputs the index together with an inference result of the hand load parameter. the estimated value of the hand load parameter obtained by the inference unit is an average value of a probability distribution updated by sequential identification from the motion data and the difference,
14. The hand load estimating device according to claim 13, wherein the index indicating the reliability of the estimated value is a variance value of the probability distribution. A hand load estimation device according to any one of claims 1 to 14,
a robot control device that generates a command to be sent to a robot based on the hand load parameters output from the hand load estimation device;
A robot control system comprising: The robot control system according to claim 15, characterized in that the robot control device adjusts the acceleration and deceleration of the robot based on the hand load parameters while the robot is in operation. The robot control system according to claim 15 or 16, characterized in that the robot control device adjusts parameters for detecting a collision of the robot based on the hand load parameters while the robot is in operation. A robot control system according to any one of claims 15 to 17, characterized in that the hand load estimation device is provided inside the robot control device. A robot control system according to any one of claims 15 to 17, characterized in that an inference unit of the hand load estimation device that infers the hand load parameters is provided inside the robot control device. A robot control system according to any one of claims 15 to 19;
a robot that operates according to commands from the robot control system;
A robot system comprising:

Images