EquiDexFlow: Contact-Grounded SE(3)-Equivariant Dexterous Grasp Generative Flows

2026-06-10 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

EquiDexFlow is an SE(3)-equivariant flow-matching model designed for dexterous grasp generation, addressing the common issue where kinematically plausible poses lack stable physical contact. This model jointly predicts wrist pose, joint angles, fingertip contacts, surface normals, and contact forces from an object point cloud. Its architecture ensures contact projection onto the object surface and forces within the Coulomb friction cone by construction, eliminating the need for loss penalties. Empirically verified over 200 rotations, EquiDexFlow demonstrates end-to-end SE(3) equivariance with wrist residuals below 0.04° and zero joint deviation. Trained on 8,100 force-closure grasps for the 16-DoF Allegro Hand, it achieves zero friction violations and the lowest wrench residual. Retargeted grasps for a 16-DoF LEAP Hand successfully complete open-loop pick-and-hold trials on physical robots, even with asymmetric objects at rotated poses.

Key takeaway

For robotics engineers developing dexterous manipulation systems, EquiDexFlow offers a robust method to generate physically stable grasps by integrating contact forces directly into the prediction. You should consider this approach to reduce downstream verification steps and improve real-world pick-and-hold success rates, especially for complex objects or varied orientations. This can streamline development and deployment of advanced robotic grasping applications.

Key insights

Jointly predicting grasp kinematics and contact forces ensures physically stable dexterous grasps.

Principles

• SE(3) equivariance improves grasp robustness.
• Projecting contacts ensures physical validity.

Method

Utilizes an SE(3)-equivariant flow-matching model to jointly predict all grasp components, including contact forces, then retargets via per-finger inverse kinematics.

In practice

• Generate stable grasps for 16-DoF Allegro Hand.
• Retarget decoded contacts to a 16-DoF LEAP Hand.

Topics

• Dexterous Grasping
• SE(3) Equivariance
• Generative Flows
• Robotic Manipulation
• Contact Mechanics
• Allegro Hand

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• Mana: Dexterous Manipulation of Articulated Tools — Mana reinterprets dexterous articulated tool manipulation as an animation problem, enabling zero-shot sim-to-real transfer via a coarse-to-fine pipeline.
• Equivariant Flow Matching for Symmetry-Breaking Bifurcation Problems — Equivariant Flow Matching accurately models multimodal distributions and symmetry-breaking bifurcations in complex dynamical systems.
• HRDexDB: A Large-Scale Dataset of Dexterous Human and Robotic Hand Grasps — HRDexDB is a multi-modal dataset for dexterous grasping, combining human and robotic hand data.
• SECOND-Grasp: Semantic Contact-guided Dexterous Grasping — Integrating semantic reasoning with physical stability enhances robotic dexterous grasping performance and adaptability.
• Beyond Binary: Sim-to-Real Dexterous Manipulation with Physics-Grounded Contact Representation — Center-of-Pressure (CoP) enables robust sim-to-real dexterous manipulation by preserving dense, physics-grounded tactile information.
• Peripheral control enabled by distributed sensing in an octopus-inspired soft robotic arm for autonomous underwater grasping — Octopus-inspired soft robots with distributed optoelectronic sensing enable adaptive, autonomous grasping in complex environments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.