Sim2Real Reinforcement Learning for Soccer skills

TL;DR

This paper introduces an improved RL approach for humanoid robot control tasks like kicking, walking, and jumping, using curriculum training and AMP, but faces challenges in sim-to-real transfer.

Contribution

It proposes a novel RL framework combining curriculum training and Adversarial Motion Priors to enhance robot control policy learning.

Findings

RL policies are more dynamic and adaptive.

Policies outperform previous methods in simulation.

Transfer to real-world remains unsuccessful.

Abstract

This thesis work presents a more efficient and effective approach to training control-related tasks for humanoid robots using Reinforcement Learning (RL). The traditional RL methods are limited in adapting to real-world environments, complexity, and natural motions, but the proposed approach overcomes these limitations by using curriculum training and Adversarial Motion Priors (AMP) technique. The results show that the developed RL policies for kicking, walking, and jumping are more dynamic, and adaptive, and outperformed previous methods. However, the transfer of the learned policy from simulation to the real world was unsuccessful, highlighting the limitations of current RL methods in fully adapting to real-world scenarios.