Research map / Image generation / Autoregressive and latent generation

DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Diffusion planning and policies · 20 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DexGrasp Anything, a diffusion-based dexterous grasp generator that injects three physical constraints (surface pulling, external/self-penetration repulsion) into training and sampling, plus an LLM-enhanced object representation, and releases a large DGA grasp dataset.

Problem

Diffusion-based dexterous grasp generation often yields hand-object penetration, insufficient contact, and low success rates because it lacks physical constraints; existing grasp datasets are also narrow in scale and diversity.

Approach

Integrate differentiable physical constraint losses into diffusion training (via the estimated clean sample h0) and into sampling via a physics-guided posterior sampler with spherical Gaussian constraint, combined with LLM-derived semantic priors for object conditioning.

Key results

• On DexGraspNet, the full method with LLM reaches 57.5% Suc.6 and 90.6% Suc.1 versus 46.9% and 79.0% for the best prior baseline UGG.
• Ablation on DexGraspNet shows physics constraints raise Suc.6 from 26.6% (baseline) to 53.6% and the LLM module adds a further gain to 57.5%.
• Training on the DGA dataset improves diversity of UGG, SceneDiffuser and GraspTTA while maintaining or improving success rates in cross-dataset evaluation.

Measured effects reported in the paper

Surprising findings

• Physics constraints applied only during training (without the physics-guided sampler) yield lower success (46.8% Suc.6) than the full method (53.6%), showing the sampler matters.

Limitations

• Physics constraints rely on hand-tuned thresholds and loss weights (dThreshold, alpha_i) not analyzed in the text.
• Sim-based evaluation in IsaacGym; real-world evaluation limited to a ShadowHand demonstration.

Experimental setup

Datasets: DexGraspNet UniDexGrasp MultiDex RealDex DexGRAB

Benchmarks: DexGraspNet UniDexGrasp MultiDex RealDex DexGRAB

Metrics: Suc. 6 Suc. 1 Pen. (mm) Div.

Baselines: UniDexGrasp GraspTTA SceneDiffuser UGG

Builds on: DDPM classifier guidance UniDexGrasp SceneDiffuser Point Transformer

Closest papers by meaning

• DexFuncGrasp: A Robotic Dexterous Functional Grasp Dataset Constructed from a Cost-Effective Real-Simulation Annotation System (AAAI 2024 · similarity 0.90)
• CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes (NeurIPS 2025 · similarity 0.90)
• Cross-Embodiment Dexterous Grasping with Reinforcement Learning (ICLR 2025 · similarity 0.88)
• UniGraspTransformer: Simplified Policy Distillation for Scalable Dexterous Robotic Grasping (CVPR 2025 · similarity 0.88)
• ZeroGrasp: Zero-Shot Shape Reconstruction Enabled Robotic Grasping (CVPR 2025 · similarity 0.88)
• ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.88)