Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dual Action Policy for Robust Sim-to-Real Reinforcement Learning

Ng Wen Zheng Terence, Chen Jianda

Abstract

This paper presents Dual Action Policy (DAP), a novel approach to address the dynamics mismatch inherent in the sim-to-real gap of reinforcement learning. DAP uses a single policy to predict two sets of actions: one for maximizing task rewards in simulation and another specifically for domain adaptation via reward adjustments. This decoupling makes it easier to maximize the overall reward in the source domain during training. Additionally, DAP incorporates uncertainty-based exploration during training to enhance agent robustness. Experimental results demonstrate DAP's effectiveness in bridging the sim-to-real gap, outperforming baselines on challenging tasks in simulation, and further improvement is achieved by incorporating uncertainty estimation.

Dual Action Policy for Robust Sim-to-Real Reinforcement Learning

Abstract

This paper presents Dual Action Policy (DAP), a novel approach to address the dynamics mismatch inherent in the sim-to-real gap of reinforcement learning. DAP uses a single policy to predict two sets of actions: one for maximizing task rewards in simulation and another specifically for domain adaptation via reward adjustments. This decoupling makes it easier to maximize the overall reward in the source domain during training. Additionally, DAP incorporates uncertainty-based exploration during training to enhance agent robustness. Experimental results demonstrate DAP's effectiveness in bridging the sim-to-real gap, outperforming baselines on challenging tasks in simulation, and further improvement is achieved by incorporating uncertainty estimation.

Paper Structure

This paper contains 17 sections, 9 equations, 3 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dynamics Mismatch in Sim-to-Real RL
  4. Uncertainty in Deep Reinforcement Learning
  5. Background
  6. Reinforcement Learning
  7. Problem Statement.
  8. Domain Adaptation with Rewards from Classifier (DARC)
  9. Methodology
  10. Dual Action Policy (DAP)
  11. Regularization
  12. Uncertainty-based Robust Action Resampling
  13. Experiments
  14. Experimental Setup
  15. Main Results
  16. ...and 2 more sections

Figures (3)

  • Figure 1: (Left) Training in Source Env: $\pi_\text{DAP}$ predicts additional action set $a^\text{tgt}$, optimized for the target environment while $a^\text{src}$ is utilized for sampling the source environment. (Right) Deployment in Target Env: only $a^\text{tgt}$ is utilized while $a^\text{src}$ is discarded.
  • Figure 2: Main results evaluated in target environment. We compare our proposed methods, DAP and DAP+U, against several baseline approaches.
  • Figure :