Researchers introduce ADEPT, a large-scale reinforcement-learning framework designed to make dexterous manipulation skills transferable from simulation to real robots. The system learns directly from raw visuo-tactile perception and targets high-degree-of-freedom embodiments solving long-horizon tasks.
ADEPT first pretrains a policy on a generic object-reposing task, then post-trains it for new downstream behaviors. The pretrained policy can zero-shot handle the reposing phase of a new task, but the authors found that naive RL fine-tuning quickly destroys that capability. Their post-training recipe combines behavior-cloning distillation, critic warm-up, and conservative on-policy updates to keep the prior intact.
The final policies are distilled into perceptive student networks that transfer zero-shot to two physical systems: a 23-DoF Kuka-Allegro hand with two RGB cameras and a 29-DoF Flexiv-Sharpa hand with two cameras and five tactile sensors. The authors report that the robots solve long-horizon tasks from challenging initial states at human-level speed. The paper is accepted at CoRL 2026.