Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2608.19182v1

Read the source

Source

  • Source name: arXiv
  • Source kind: paper
  • Authors / organization: Jayjun Lee, Jessica Yin, Asif Rana, Nicholas Blauch, Sam Mady, Mohak Bhardwaj, Nima Fazeli, Nathan Ratliff, Karl Van Wyk, Ankur Handa
  • Venue: arXiv: cs.RO
  • Published: 2026-08-19
  • Collected: 2026-08-20
  • URL: http://arxiv.org/abs/2608.19182v1
  • PDF: https://arxiv.org/pdf/2608.19182v1
  • DOI: UNKNOWN
  • Registry group: academic_apis_and_preprints
  • Evidence role: canonical_academic_index
  • Language: en
  • Country / region: UNKNOWN

Plain-Language Brief

The paper introduces ADEPT, a reinforcement learning framework for training multi-fingered robots to perform complex tasks directly from raw visual and tactile perception. The framework involves pre-training a dexterous policy on a generic object reposing task, then post-training downstream policies using this pretrained behavior as a prior. A stable post-training recipe is used to maintain the pretrained policy's capabilities during transfer. A joint-space Geometric Fabric is introduced to safely exploit the robot's full kinematic dexterity. The paper demonstrates zero-shot sim-to-real transfer on two embodiments: a 23 DoF Kuka-Allegro and a 29 DoF Flexiv-Sharpa.

What the Document Says

  • The ADEPT framework is introduced for reinforcement learning.
  • ADEPT is designed to learn sim-to-real transferable dexterity across high degree-of-freedom robot embodiments.
  • Pre-training a dexterous policy on a generic object reposing task is part of the ADEPT framework.
  • Post-training downstream policies with this pretrained behavior as a prior is part of the ADEPT framework.
  • A stable post-training recipe is used to maintain the pretrained policy's capabilities during transfer.
  • Behavior-cloning distillation, critic warm-up, and conservative on-policy updates are part of the stable post-training recipe.
  • A joint-space Geometric Fabric is introduced in the ADEPT framework to safely exploit the full kinematic dexterity.
  • The paper demonstrates zero-shot sim-to-real transfer on two embodiments: a 23 DoF Kuka-Allegro and a 29 DoF Flexiv-Sharpa.

Key Concepts Mentioned

  • ADEPT framework
  • Reinforcement learning
  • Sim-to-real transfer
  • Dexterity
  • High degree-of-freedom robot embodiments
  • Pre-training
  • Post-training
  • Stable post-training recipe
  • Behavior-cloning distillation
  • Critic warm-up
  • Conservative on-policy updates
  • Joint-space Geometric Fabric

Claims Made by the Document

  • ADEPT can solve long-horizon tasks directly from raw visuo-tactile perception.
  • ADEPT enables learning new behaviors that are otherwise difficult to discover from scratch on multi-fingered robots.
  • ADEPT 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.
  • The introduced joint-space Geometric Fabric safely exploits the full kinematic dexterity.

Evidence Used

  • The text describes the ADEPT framework and its components.
  • The text explains how ADEPT is used to solve long-horizon tasks and learn new behaviors.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.