IROS 2026 DexKP VLA - Heungwoo/research GitHub Wiki

IROS 2026 — DexKP-VLA: Dexterous Keypoint VLA for Zero-Shot Grasping

Venue: IROS 2026 (Pittsburgh) · paper #752 · Huazhong Univ. of S&T · Wuhan University (Hu, Yang, Xiong). Representative of: dexterous grasping via VLA — a hierarchical keypoint bridge that lets a VLM drive a five-finger hand zero-shot, no fine-tuning. Companions: Dexterous Manipulation · Dexterous-Hand Data Pyramid · VLA Architectures · IROS 2026 survey.

1. Problem

VLAs revolutionized manipulation but are data- and compute-hungry and largely restricted to parallel-jaw grippers — not five-finger dexterous hands. Can a pretrained VLM drive dexterous grasping zero-shot without the usual data/compute cost?

2. Method

DexKP-VLA decouples high-level semantic planning from low-level geometric control via keypoints:

  • A pretrained VLM decomposes the task (semantic planning).
  • A Keypoint Construction Module (KCM) bridges the cross-modal gap by grounding abstract semantic goals into sparse, actionable geometric keypoints.
  • The keypoints guide a physics-constrained optimization solver to generate precise trajectories — so the dexterous control is optimization, not a learned action head, avoiding the data cost.

3. Results

Real-world, 40+ unseen objects, no task-specific fine-tuning:

  • Single-object grasping 81.9% · cluttered grasping 76.8% · pick-and-place 72.6% (category-averaged).

4. Why it matters (dexterous lens)

DexKP-VLA is the IROS 2026 representative of the "VLM semantics + geometric keypoints + optimization" route to dexterous grasping — a training-light alternative to data-hungry dexterous VLAs. Its keypoint interface is the transferable idea: keypoints are an embodiment-friendlier action abstraction than joint-space, connecting to the data-pyramid's retargeting concern (keypoints sidestep per-hand action learning). It pairs with VCoT-Grasp (language-driven grasp generation) as the two IROS grasping-foundation angles, and with DexGrasp-Zero/One-Hand on the zero-shot-to-unseen-hands thread.

Limitations (reviewer): relies on a physics-constrained solver (needs object geometry/pose) — a modular pipeline, not end-to-end; keypoint grounding quality bounds precision; grasping-centric (not long-horizon manipulation).

5. Links

← Back to IROS 2026 survey · Home