Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Predictive Learning: Motion Learning Concept inspired by Cognitive Robotics

Kanata Suzuki, Hiroshi Ito, Tatsuro Yamada, Kei Kase, Tetsuya Ogata

TL;DR

This paper describes the proposed concept, its implementation, and examples of its applications in real robots, based on the fundamental strategy of predicting the near-future sensorimotor states of robots and online minimization of the prediction error between the real world and the model.

Abstract

Bridging the gap between motion models and reality is crucial by using limited data to deploy robots in the real world. Deep learning is expected to be generalized to diverse situations while reducing feature design costs through end-to-end learning for environmental recognition and motion generation. However, data collection for model training is costly, and time and human resources are essential for robot trial-and-error with physical contact. We propose "Deep Predictive Learning," a motion learning concept that predicts the robot's sensorimotor dynamics, assuming imperfections in the prediction model. The predictive coding theory inspires this concept to solve the above problems. It is based on the fundamental strategy of predicting the near-future sensorimotor states of robots and online minimization of the prediction error between the real world and the model. Based on the acquired sensor information, the robot can adjust its behavior in real time, thereby tolerating the difference between the learning experience and reality. Additionally, the robot was expected to perform a wide range of tasks by combining the motion dynamics embedded in the model. This paper describes the proposed concept, its implementation, and examples of its applications in real robots. The code and documents are available at: https://ogata-lab.github.io/eipl-docs

Deep Predictive Learning: Motion Learning Concept inspired by Cognitive Robotics

TL;DR

This paper describes the proposed concept, its implementation, and examples of its applications in real robots, based on the fundamental strategy of predicting the near-future sensorimotor states of robots and online minimization of the prediction error between the real world and the model.

Abstract

Bridging the gap between motion models and reality is crucial by using limited data to deploy robots in the real world. Deep learning is expected to be generalized to diverse situations while reducing feature design costs through end-to-end learning for environmental recognition and motion generation. However, data collection for model training is costly, and time and human resources are essential for robot trial-and-error with physical contact. We propose "Deep Predictive Learning," a motion learning concept that predicts the robot's sensorimotor dynamics, assuming imperfections in the prediction model. The predictive coding theory inspires this concept to solve the above problems. It is based on the fundamental strategy of predicting the near-future sensorimotor states of robots and online minimization of the prediction error between the real world and the model. Based on the acquired sensor information, the robot can adjust its behavior in real time, thereby tolerating the difference between the learning experience and reality. Additionally, the robot was expected to perform a wide range of tasks by combining the motion dynamics embedded in the model. This paper describes the proposed concept, its implementation, and examples of its applications in real robots. The code and documents are available at: https://ogata-lab.github.io/eipl-docs
Paper Structure (27 sections, 1 equation, 7 figures)

This paper contains 27 sections, 1 equation, 7 figures.

Table of Contents

  1. Introduction
  2. Experience-based Robotics
  3. Proposed Robot Learning Concept
  4. Deep Learning in Robotics Fields
  5. Recognition
  6. Robot Motion Learning
  7. Reinforcement Learning
  8. Behavioral Cloning
  9. Deep Predictive Learning
  10. Main Concept and Policies
  11. Implementation
  12. Learning Sensorimotor Dynamics
  13. Robot Motion Generation based on Minimizing Prediction Errors
  14. Hierarchical Structure with Different Time-scales
  15. Bottom-up and Top-down Attention
  16. ...and 12 more sections

Figures (7)

  • Figure 1: Overview of the free energy principle Friston2010nrn.
  • Figure 2: An example of a deep predictive learning framework consisting of a feature extraction unit and a time-series prediction unit.
  • Figure 3: Real-time switching of multiple motion generation models based on prediction errors Ito2022sr.
  • Figure 4: Embedding attractor dynamics into the internal state of RNN Kase2018icra.
  • Figure 5: Grounding of language instructions and motion generation in the designed internal dynamics of RNN Yamada2016fn.
  • ...and 2 more figures