The authors derived Soft MellowMax Actor-Critic (SMM-AC), extending soft mellowmax value functions to continuous state-action spaces by approximating the intractable reference policy with a slowly updated reference network.
Regularised reinforcement learning boosts policy robustness in sim-to-real transfer
A novel algorithm based on soft mellowmax achieves higher real-world returns and greater resilience to perturbations than standard methods.
Big Tech
Danilo Vucetic · Mohamad H. Danesh · Florence Clerc · Hsiu-Chin Lin · Jonathan Tremblay
Mila - Québec AI Institute · Polytechnique Montréal · McGill University · NVIDIA · Université de Montréal
Research Digest··2 min read
The authors propose SMM-AC, a regularised RL algorithm extending soft mellowmax to continuous control, and evaluate it against PPO, SAC, and TD3 on MuJoCo benchmarks and real robots (Franka arm, ANYmal-D).
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.