Recoleta
GitHub
Source note

Dexora: Open-source VLA for High-DoF Bimanual Dexterity

Authors
Zongzheng Zhang, Jingrui Pang, Zhuo Yang, Kun Li, Minwen Liao, Saining Zhang, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionDexterous ManipulationBimanual RoboticsRobot Data ScalingDiffusion Policy
Open arXivMarkdown source

Summary

Dexora is an open-source vision-language-action system for dual-arm, dual-hand high-DoF robot manipulation. It combines a 36-DoF robot, matched simulation, large real and synthetic datasets, and quality-weighted diffusion policy training.

Problem

  • Existing VLA systems mainly target dual-arm grippers or single-arm dexterous hands, so they miss tasks that need two arms and articulated fingers.
  • The gap matters for tasks such as bottle-cap twisting, pen use, cutting food, fetching books from tight shelves, and bimanual object separation.
  • Real teleoperation data for high-DoF hands is noisy because of operator variation, tracking errors, occlusion, and latency.

Approach

  • The hardware uses two 6-DoF AIRBOT arms and two 12-DoF XHAND dexterous hands, giving 36 DoF in total.
  • Teleoperation splits the control problem: a custom exoskeleton backpack records arm motion, while Apple Vision Pro hand tracking records finger motion.
  • The same teleoperation interface drives the real robot and a MuJoCo digital twin, with four RGB views and full joint states logged at 20 Hz.
  • The training corpus combines 100K simulated trajectories with 6.5M frames and 10K real teleoperated episodes with 2.92M frames.
  • The policy is a diffusion transformer conditioned on language, multi-view images, and joint state; an offline discriminator scores clip quality and weights the diffusion loss so low-quality demonstrations count less during training.

Results

  • On 12 basic real-world tasks, Dexora reports an 89.6% average success rate, compared with 82.1% for GR00T N1, 50.4% for π0, and 34.2% for Diffusion Policy.
  • Dexora reaches at least 90% success on 7 of 12 basic tasks and scores 100% on apple-to-plate, bowl-to-bowl, and cabinet-door opening.
  • On 6 dexterous tasks, Dexora reports 66.7% average success, compared with 51.7% for GR00T N1, 26.7% for π0, and 6.7% for Diffusion Policy.
  • Dexterous task gains include 65% on Use pen, 80% on Fetch book, 80% on Rough dough, and 25% on Twist cap; all baselines score 0% on Twist cap.
  • The paper claims cross-embodiment transfer to single-arm grippers, dual-arm grippers, and single-arm low-DoF hands without changing the model architecture, but the excerpt does not provide detailed success-rate tables for those transfer tests.

Link