residual reinforcement learning

Residual reinforcement learning is the focus of research describing a practical way to improve robot behavior without retraining an entire vision-language-action model. Microsoft Research’s object-centric method learns a lightweight corrective policy in simulation, then combines it with a frozen model for real-robot execution. The tagged discussion centers on the gap between capable demonstrations and reliable manipulation, rather than presenting a general solution to household robotics or production deployment. Reported results cover five manipulation tasks, with average success improving from 42 percent to 76 percent in zero-shot real-robot tests. This tag is useful for readers following embodied AI, simulation-based robot training, and methods for adding targeted reflexes to existing models.