Asymmetric Actor-Critic for Image-Based Robot Learning

AI Impact Summary

This CAPABILITY introduces an asymmetric actor-critic approach for training policies from image inputs in robotics. By giving the critic access to privileged state information while the actor learns from images, it aims to improve sample efficiency and policy stability in vision-based manipulation tasks. Adoption would impact robotics RL pipelines, requiring alignment between perception outputs, data collection, and training code to exploit the asymmetry.