Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Loss Function Considering Dead Zone for Neural Networks

Koki Inami, Koki Yamane, Sho Sakaino

TL;DR

This work tackles the limited training data problem caused by actuator dead zones in learning inverse dynamics for manipulators. It introduces a dead-zone aware loss that ignores errors from joints in dead zones by employing a per-joint reliability indicator and a loss function $L$ that sums only over non-dead-zone joints. Empirical results on a 3-DOF manipulator show improved accuracy compared with Newton–Euler and a conventional NN, with notable gains in data efficiency (approximately using $70\%$ of data versus $34\%$ for the baseline). The method holds promise for extending NN-based inverse dynamics to higher-DOF systems and other learning models, enabling more robust model-based control in the presence of dead zones.

Abstract

It is important to reveal the inverse dynamics of manipulators to improve control performance of model-based control. Neural networks (NNs) are promising techniques to represent complicated inverse dynamics while they require a large amount of motion data. However, motion data in dead zones of actuators is not suitable for training models decreasing the number of useful training data. In this study, based on the fact that the manipulator joint does not work irrespective of input torque in dead zones, we propose a new loss function that considers only errors of joints not in dead zones. The proposed method enables to increase in the amount of motion data available for training and the accuracy of the inverse dynamics computation. Experiments on actual equipment using a three-degree-of-freedom (DOF) manipulator showed higher accuracy than conventional methods. We also confirmed and discussed the behavior of the model of the proposed method in dead zones.

Loss Function Considering Dead Zone for Neural Networks

TL;DR

This work tackles the limited training data problem caused by actuator dead zones in learning inverse dynamics for manipulators. It introduces a dead-zone aware loss that ignores errors from joints in dead zones by employing a per-joint reliability indicator and a loss function that sums only over non-dead-zone joints. Empirical results on a 3-DOF manipulator show improved accuracy compared with Newton–Euler and a conventional NN, with notable gains in data efficiency (approximately using of data versus for the baseline). The method holds promise for extending NN-based inverse dynamics to higher-DOF systems and other learning models, enabling more robust model-based control in the presence of dead zones.

Abstract

It is important to reveal the inverse dynamics of manipulators to improve control performance of model-based control. Neural networks (NNs) are promising techniques to represent complicated inverse dynamics while they require a large amount of motion data. However, motion data in dead zones of actuators is not suitable for training models decreasing the number of useful training data. In this study, based on the fact that the manipulator joint does not work irrespective of input torque in dead zones, we propose a new loss function that considers only errors of joints not in dead zones. The proposed method enables to increase in the amount of motion data available for training and the accuracy of the inverse dynamics computation. Experiments on actual equipment using a three-degree-of-freedom (DOF) manipulator showed higher accuracy than conventional methods. We also confirmed and discussed the behavior of the model of the proposed method in dead zones.
Paper Structure (6 sections, 4 equations, 6 figures)

This paper contains 6 sections, 4 equations, 6 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Experimental Design
  5. Result
  6. Conclusion

Figures (6)

  • Figure 1: Backpropagation in the conventional method and the proposed method
  • Figure 4: CRANE-X7 (RT)
  • Figure 5: Block diagram of 4-channel bilateral control
  • Figure 6: Block diagram of controller
  • Figure 7: The average MSE and 95 percent confidence interval on the test data obtained by 10 trials in which the specified number of data are randomly selected from 6000 samples of the training data are shown.
  • ...and 1 more figures