Event
ADEPT: Pretraining and Post-Training Reinforcement Learning Accelerates Dexterous Manipulation
1 reports · 1 sources
Jayjun Lee et al. propose ADEPT, a large-scale reinforcement learning framework that first pretrains dexterous policies on general object rearrangement tasks, then uses this prior to post-train downstream policies for sim-to-real dexterous manipulation learning on high-DOF robot embodiments, capable of solving long-horizon tasks directly from raw visual-tactile perception. To avoid naive fine-tuning destroying zero-shot rearrangement capability, the authors propose a stable post-training scheme combining behavior cloning distillation, critic warm-up, and conservative on-policy updates, and introduce a joint-space Geometric Fabric coordination policy with the robot. The authors distill the post-trained teacher into a perception student, achieving zero-shot sim-to-real transfer, validated on two embodiments: a 23-DOF Kuka-Allegro and a 29-DOF Flexiv-Sharpa. The paper was accepted to CoRL 2026.
Report timeline
- ADEPT: Accelerating Dexterous Manipulation via Reinforcement Learning Pre-Training and Post-Training
ADEPT is a large-scale reinforcement learning framework that first pre-trains a dexterous policy on a generic object-reposing task, then post-trains downstream policies with this prior, enabling long-horizon tasks to be solved directly from raw visual-tactile perception across high-DoF embodiments. To prevent naive fine-tuning from degrading zero-shot reposing ability, the authors propose a stable post-training scheme combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates, and introduce a joint-space Geometric Fabric that coordinates the policy with the robot.
arXiv cs.RO ·