Event
SMM-AC: Regularized Reinforcement Learning Improves Real-Robot Transfer
1 reports · 1 sources
Danilo Vucetic and co-authors propose SMM-AC, a regularized reinforcement learning algorithm based on a soft mellowmax value function, and validate it on MuJoCo continuous-control benchmarks and on a real robot. The authors report that on the MuJoCo benchmarks SMM-AC performs comparably to baselines such as SAC; on the real robot, SMM-AC achieves higher returns and is more robust than the baselines to perturbations in observations, the environment, and dynamics. The study compares the sim-to-real transfer performance of commonly used reinforcement learning algorithms, concluding that regularization can improve the robustness of policies to the sim-to-real gap and to real-environment variation.
Report timeline
- SMM-AC: Regularized reinforcement learning improves sim-to-real transfer
The authors propose SMM-AC, a regularized reinforcement learning algorithm based on the soft mellowmax value function. On MuJoCo continuous control benchmarks it is comparable to baselines such as SAC, and on real robots it achieves higher returns while being more robust to perturbations in observations, the environment, and dynamics. The study compares the sim-to-real transfer performance of commonly used reinforcement learning algorithms and finds that regularization improves policy robustness to the sim-to-real gap and to real-world environment changes.
arXiv cs.RO ·