SECOND-Grasp: Semantic Contact-guided Dexterous Grasping

2026-05-13 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

SECOND-Grasp, a unified framework for robotic dexterous grasping, integrates physical stability with semantic task guidance. The system first uses vision-language reasoning to generate coarse contact proposals based on object properties, localizing these regions across multiple views. It then employs Semantic-Geometric Consistency Refinement (SGCR) to ensure cross-view consistency and remove geometrically invalid regions, producing reliable 3D contact maps. From these maps, feasible hand poses are derived via inverse kinematics, which serve as supervision for policy learning. Trained on DexGraspNet, SECOND-Grasp achieved 98.2% lifting success on seen categories and 97.7% on unseen categories, outperforming baselines. It also improved intent-aware grasping by 12.8% and 26.2% and showed promising results with Shadow Hand and Allegro Hand.

Key takeaway

For research scientists developing advanced robotic manipulation systems, SECOND-Grasp demonstrates that integrating semantic understanding with physical stability is crucial for robust dexterous grasping. You should consider incorporating vision-language reasoning and geometric consistency refinement into your grasp generation pipelines to improve both lifting success rates and intent-aware manipulation across diverse objects and robotic hands.

Key insights

Integrating semantic reasoning with physical stability enhances robotic dexterous grasping performance and adaptability.

Principles

• Combine vision-language reasoning with geometric validation.
• Refine initial predictions for cross-view consistency.

Method

SECOND-Grasp uses vision-language reasoning for coarse contact proposals, refines them with Semantic-Geometric Consistency Refinement (SGCR) for 3D contact maps, and derives hand poses via inverse kinematics for policy learning.

In practice

• Apply SGCR for robust 3D contact map generation.
• Use inverse kinematics to generate grasp supervision signals.

Topics

• Dexterous Grasping
• Semantic Reasoning
• Vision-Language Reasoning
• 3D Contact Maps
• Grasp Policy Learning

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

Related on AIssential

• Bridging the Morphology Gap: Adapting VLA Models to Dexterous Manipulation via Intent-Conditioned Fine-Tuning — Adapting VLA models to dexterous hands is possible by decoupling control via a virtual grasp intent and a two-stage learning architecture.
• DragMesh-2: Physically Plausible Dexterous Hand-Object Interaction with Articulated Objects — DragMesh-2 enables robust dexterous interaction with articulated objects by modeling contact dynamics and injecting physical signals into policy learning.
• NoContactNoWorries: Estimating Contact through Vision and Proprioception for In-Hand Dexterous Manipulation — Robots can infer binary contact states for dexterous manipulation by fusing RGB-D vision and proprioception, offering a scalable alternative to hardware tactile sensors.
• Pro$^2$Assist: Continuous Step-Aware Proactive Assistance with Multimodal Egocentric Perception for Long-Horizon Procedural Tasks — Pro²Assist offers continuous, step-aware proactive assistance for complex tasks using multimodal egocentric perception from AR glasses.
• Blind Dexterous Grasping via Real2Sim2Real Tactile Policy Learning — Calibrated simulation and geometry-aware tactile representations enable blind dexterous grasping on physical robots.
• EquiDexFlow: Contact-Grounded SE(3)-Equivariant Dexterous Grasp Generative Flows — Jointly predicting grasp kinematics and contact forces ensures physically stable dexterous grasps.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.