Robotics

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

Zongzheng Zhang, Jingrui Pang, et al.

Featured May 20, 2026

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new robot system called Dexora learns to use two hands with many fingers by watching humans, but it cleverly ignores shaky human movements to learn super smooth and tricky skills, even for different robots.

In depth

The paper introduces Dexora, the first open-source Vision-Language-Action (VLA) system designed for dual-arm, dual-hand, high-degree-of-freedom (DoF) manipulation. It achieves this by combining a novel hybrid teleoperation pipeline for data collection with a data-quality-aware training recipe. This recipe uses an offline discriminator to assign clip-level weights to demonstrations, effectively down-weighting low-quality data during diffusion-transformer policy training, leading to superior dexterity and generalization.

Key Takeaways

1
Dexora is the first VLA system to natively support dual-arm, dual-hand, high-DoF manipulation, addressing a critical gap in embodied AI.
2
A hybrid teleoperation pipeline (exoskeleton for arms, Apple Vision Pro for fingers) and a large, embodiment-matched dataset (100K simulated, 10K real trajectories) enable scalable and high-fidelity data collection.
3
A discriminator-guided quality-aware training approach mitigates noisy teleoperation data by weighting demonstrations based on kinematic smoothness, task completion, and policy compatibility, significantly improving policy learning for dexterous skills.

Conceptual Flow

HIGH LEVEL

Methodology: Smart Data & Training

The system collects lots of robot movements, both real and fake, then uses a smart filter to only learn from the best, smoothest actions.

Human Control

Robot & Twin

Collect Data

Mixed Quality Actions

Visuals & Language

Results: Dexterous & Adaptable Robots

This new way of learning helps robots do complex tasks with two hands much better and even lets them use the same skills on simpler robots.

Old Robot Skills

New Robot Skills

Compare Performance

Better Dexterity

Works on Many Robots