Venue: CoRL 2025 (Oral, ~5% of submissions) · Authors: Zeyuan Chen, Qiyang Yan, Yuanpei Chen, Tianhao Wu, Jiyao Zhang, Zihan Ding, Jinzhou Li, Yaodong Yang, Hao Dong (corresponding) — Peking University (CFCS, PKU-AgiBot Lab, PKU-PsiBot Lab) + Princeton (Z. Ding) · arXiv: 2506.14317 · OpenReview: 4XKKUifQ9c Category: Dexterous Manipulation Trend tag: Closed-loop dex at clutter scale

flowchart LR
  SIM[Large-scale sim<br/>cluttered scenes + dex hands] --> POL[Dex grasping policy<br/>closed-loop, target-oriented]
  POL -- zero-shot --> REAL[Real dex grasping<br/>in clutter]

Dexterous grasping in clutter is the natural frontier after single-object grasping: you have to reason about collisions, re-orient, and occasionally push other objects. Open-loop planners fail, and closed-loop learned policies had not been shown to transfer zero-shot from simulation.

A two-stage teacher-student framework for closed-loop, target-oriented dexterous grasping:

• Teacher is trained in simulation via RL with a clutter-density curriculum, using a geometry- and spatially-embedded scene representation and a novel comprehensive safety curriculum to learn collision-aware, safe grasping behaviors.
• The teacher is then distilled by imitation learning into a student 3D diffusion policy (DP3) that operates on partial point-cloud observations, enabling zero-shot deployment on real hardware in cluttered scenes.

CoRL 2025 Oral (~5% of submissions). Claims the first zero-shot sim-to-real closed-loop system for target-oriented dexterous grasping in cluttered scenes. Real-world evaluation spans 41 objects of diverse shapes/sizes/materials across 9 sparse, 5 dense, and 3 ultra-dense scenes, plus a continuous grasp-and-transport demo over 30 unique objects (run until 3 consecutive failures).

Sets a new bar for what "dexterous manipulation in the wild" means at CoRL. Part of the CoRL 2025 dex wave alongside DexUMI, DexSkin, KineSoft. Threads into ICLR 2026's DexNDM, UniHM, RFS.

• arXiv: https://arxiv.org/abs/2506.14317

• Survey (CoRL 2025)

← Back to CoRL-2025