---
title: "ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.19182v1"
permalink: "https://feed.astralor.com/items/itm_0bc4d907efc2fb92cbf934d3"
published: "2026-08-19T17:55:06.000Z"
score: 78
channel: "ai"
tags: ["ADEPT","Flexiv-Sharpa","Kuka-Allegro","Sim-to-Real Transfer","Reinforcement Learning","Dexterous Manipulation"]
---

# ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning

## AI commentary

> This work transfers the pre-training paradigm to dexterous manipulation. The most notable contribution is solving capability degradation during naive fine-tuning and achieving zero-shot sim-to-real transfer on multi-fingered hands with visuo-tactile inputs. For researchers in robot learning, RL algorithms, or embodied AI, ADEPT's stable post-training recipe and Geometric Fabric design offer direct reference value.

## At a glance

ADEPT is a large-scale reinforcement learning framework designed to accelerate sim-to-real transferable dexterity across high degree-of-freedom robot embodiments via pre-training and post-training. It first pre-trains a dexterous policy on a generic object reposing task, then uses this pretrained behavior as a prior for post-training downstream policies, avoiding learning from scratch. The framework introduces a stable post-training recipe combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates, along with a joint-space Geometric Fabric. It demonstrates zero-shot sim-to-real transfer on a 23-DoF Kuka-Allegro and a 29-DoF Flexiv-Sharpa, solving long-horizon tasks at human-level speed.

## Key points

- Proposes ADEPT, a pre-training + post-training framework that accelerates dexterous manipulation learning on multi-fingered robots, avoiding re-learning basic skills for each new task.
- Introduces a stable post-training recipe (behavior-cloning distillation, critic warm-up, conservative on-policy updates) to mitigate capability degradation during transfer.
- Presents a joint-space Geometric Fabric that mediates between RL policy and robot, enabling safe and full exploitation of kinematic dexterity.
- Achieves zero-shot sim-to-real transfer on 23-DoF Kuka-Allegro and 29-DoF Flexiv-Sharpa, solving long-horizon tasks from challenging initial states at human-level speed.

## Full text

We introduce Accelerating Dexterity via Pre-Training (ADEPT), a large-scale reinforcement learning (RL) framework for learning sim-to-real transferable dexterity across high degree-of-freedom (DoF) robot embodiments that can solve long-horizon tasks directly from raw visuo-tactile perception. ADEPT pretrains a dexterous policy on a generic object reposing task, then post-trains downstream policies with this pretrained behavior as a prior. ADEPT enables learning new behaviors that are otherwise difficult to discover from scratch on multi-fingered robots and avoids learning the same set of skills over again for every new downstream task. The pretrained policy zero-shots the reposing phase of downstream tasks, but naïve RL fine-tuning rapidly degrades this capability during transfer. We address this with a stable post-training recipe combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates. To safely exploit the full kinematic dexterity, we introduce a joint-space Geometric Fabric that mediates between the RL policy and the robot. We distill post-trained teachers into perceptive students that zero-shot sim-to-real transfer on two embodiments: a 23 DoF Kuka-Allegro with two RGB cameras, and a 29 DoF Flexiv-Sharpa with two RGB cameras and five vision-based tactile sensors, and can solve long-horizon tasks from challenging initial states with dexterity at human-level speed.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_0bc4d907efc2fb92cbf934d3