ADEPT: Sim-to-Real Dexterous Manipulation via RL Pre-Training and Post-Training

Forum topic · 小凯 · 2026-08-21

Summary

ADEPT (Accelerating Dexterity via Pre-Training) is a large-scale reinforcement learning framework from researchers including Jayjun Lee and Ankur Handa for learning sim-to-real transferable dexterity on high degree-of-freedom robot embodiments, solving long-horizon tasks directly from raw visuo-tactile perception. The framework pretrains a dexterous policy on a generic object reposing task, then post-trains downstream policies using this pretrained behavior as a prior. This enables learning new behaviors that are difficult to discover from scratch on multi-fingered robots and avoids relearning the same skills for each new task. Because naive RL fine-tuning degrades the pretrained reset capability, the authors combine behavior-cloning distillation, critic warm-up, and conservative on-policy updates for stable post-training, plus a joint-space geometric fabric mediating between the RL policy and the robot. The final teacher is distilled into a perception-based student achieving zero-shot sim-to-real transfer on a 23-DoF Kuka-Allegro (two RGB cameras) and a 29-DoF Flexiv-Sharpa (two RGB cameras plus five vision-based tactile sensors), solving long-horizon tasks at human-level speed from challenging initial states. Paper: arXiv 2608.19182.

论文概要

Field: Machine Learning Authors: Jayjun Lee, Jessica Yin, Asif Rana, Nicholas Blauch, Sam Mady, Mohak Bhardwaj, Nima Fazeli, Nathan Ratliff, Karl Van Wyk, Ankur Handa Published: 2026-08-19 arXiv: 2608.19182

Introduction

ADEPT (Accelerating Dexterity via Pre-Training) is a large-scale reinforcement learning (RL) framework for learning sim-to-real transferable dexterity across high degree-of-freedom (DoF) robot embodiments, capable of solving long-horizon tasks directly from raw visuo-tactile perception.

Key Ideas

Pre-training on object reposing: ADEPT pretrains a dexterous policy on a generic object reposing task, then uses this pretrained behavior as a prior to post-train downstream policies.
Enables hard-to-discover behaviors: The prior makes it possible to learn new behaviors that would otherwise be difficult to discover from scratch on multi-fingered robots, and avoids relearning the same skills for every downstream task.
Zero-shot reset handling: The pretrained policy zero-shots the reposing (reset) phase of downstream tasks.

Stable Post-Training

Naive RL fine-tuning rapidly degrades the pretrained capability during transfer. The authors address this with:

Behavior cloning distillation
Critic warm-up
Conservative on-policy updates

To safely exploit full kinematic dexterity, they introduce a joint-space geometric fabric that mediates between the RL policy and the robot.

Sim-to-Real Results

The post-trained teacher is distilled into a perception-based student, achieving zero-shot sim-to-real transfer on two embodiments:

23-DoF Kuka-Allegro with two RGB cameras
29-DoF Flexiv-Sharpa with two RGB cameras and five vision-based tactile sensors

Both solve long-horizon tasks at human-level speed from challenging initial states.