Reproduce “Risk-Aware Reinforcement Learning for Mobile Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Michael Groom, James Wilson, Nick Hawes, Lars Kunze
Researchers from the Oxford Robotics Institute developed the first framework that integrates Distributional Reinforcement Learning with distortion risk metrics to train risk-aware visuomotor policies for mobile manipulation. This framework enables runtime-adjustable risk sensitivity and successfully transfers these behaviors from privileged teacher policies to vision-based student policies via imitation learning, achieving comparable task performance to risk-neutral baselines while improving worst-case outcomes for risk-averse settings.
Key insights
- Dynamic, risk-aware DRL policies can be effectively learned for mobile manipulation and successfully transferred to vision-based student policies through imitation learning.
- Training across a range of risk sensitivities (multi-beta training) can lead to more generalized and robust policies, even when evaluated in a risk-neutral context.
- The teacher-student distillation approach is crucial for overcoming the sample inefficiency and high VRAM costs associated with direct vision-based RL on high-dimensional visual inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.