本文由 AI 分析生成
建立時間： 2026-07-30 來源： https://arxiv.org/html/2504.18829v2
Summary
Dexonomy proposes an efficient two-stage pipeline for synthesizing contact-rich, physically plausible dexterous grasps covering all 31 types in the GRASP taxonomy. Starting from a single human-annotated template per grasp type, the pipeline first optimizes object pose to globally fit the hand template, then locally refines the hand in MuJoCo simulation. The resulting 9.5M-grasp dataset over 10.7k objects enables a type-conditional generative model that achieves 82.3% real-world success across 12 grasp types.
Dexonomy 提出一個高效的兩階段抓取合成管線，涵蓋 GRASP 分類中全部 31 種抓取類型，每種手型僅需一個人工標注模板。管線先以全域對齊優化物體姿態，再以 MuJoCo 模擬進行局部手部細化，最終資料集含 10.7k 個物體與 950 萬個抓取。以此訓練的類型條件生成模型在真實世界實驗中達 82.3% 成功率。
Prerequisites
- GRASP taxonomy — The 31-type grasp taxonomy is the target coverage requirement; understanding what it classifies (power vs. precision grasps, fingertip vs. palm contact patterns) is needed to interpret why full-taxonomy synthesis is a meaningful contribution beyond type-agnostic grasping.
- Dexterous hand kinematics and contact mechanics — The optimization pipeline operates in hand configuration space using transposed Jacobians; understanding multi-DOF hand parameterization is required to evaluate the contact-aware control strategy.
- MuJoCo physics simulation — The local refinement stage uses contact dynamics in MuJoCo to validate physically plausible grasps; familiarity with rigid-body simulation informs how penetration-free grasps are enforced.
Core Idea
The central insight is that human knowledge about grasp morphology—encoded in a single annotated template per grasp type—can be scaled to millions of synthesized grasps through a two-stage decoupling strategy. Rather than jointly optimizing hand pose and object placement simultaneously (a combinatorially complex problem), Dexonomy first solves a simpler global alignment problem (fit the object rigidly to the static hand template), then solves a second local conformance problem (move fingers to conform to the now-nearby object surface using physics simulation). This decoupling converts an intractable joint optimization into two tractable sequential ones. A contact-aware control strategy—applying force at each contact point proportional to the desired normal—converts static pose candidates into validated, functionally stable grasps that can themselves be used as new templates for bootstrapping future synthesis.
Results
| Benchmark | Dexonomy | BODex (baseline) | Delta |
|---|---|---|---|
| Grasp Success Rate (simulation) | 60.50% | 49.23% | +11.27 pp |
| Penetration Depth | 0.00 mm | 0.63 mm | −0.63 mm |
| Contact Link Number | 4.38 | 3.85 | +0.53 |
| Generative model GSR | 63.9% | 54.0% | +9.9 pp |
| Real-world success rate | 82.3% | — | — |
Real-world: Shadow hand, 13 objects, 12 grasp types, ~200 trials.
Limitations
- Author-stated: Single robotic hand platform (Shadow hand); templates require one human annotation per grasp type per hand model; real-world experiments limited to 13 objects and 12 grasp types.
- Unstated: Pipeline depends on high-quality 3D meshes (Objaverse); sim-to-real transfer is evaluated for one hand only and likely varies across hardware; type-conditional model uses single-view point clouds which may fail under occlusion; generalization to soft or deformable objects is untested.
Reproducibility
- Code: Project page at https://pku-epic.github.io/Dexonomy (check for release)
- Datasets: DexGraspNet + Objaverse (existing benchmarks); synthesized 9.5M grasp dataset release status unclear at time of writing
- Compute: GPU-parallel global alignment; MuJoCo simulation; Shadow hand hardware required for real-world replication
Insights
The grasp-type-conditioned generation framing marks a significant shift from type-agnostic grasping literature. Most prior work treats grasping as binary (success vs. failure) rather than as a structured space of interaction primitives—yet actual manipulation tasks require specific grasp types (cutting needs a power grasp; picking up a small object needs a precision pinch). The one-annotation-bootstraps-millions approach also demonstrates that minimal human labeling can scale to large datasets when the right inductive bias is built into the synthesis pipeline. The validated grasps doubling as new templates creates a compounding data flywheel.
Connections
- Learning from Demonstration — Single human-annotated template bootstrapping millions of grasps is analogous to one-shot LfD; the pipeline amplifies sparse human labels into a full dataset
- VLA models — Type-conditional grasp generation is a natural interface for VLA action heads; connecting language-described tasks to GRASP taxonomy types is a near-term integration path
- dexterous-manipulation
- grasp-synthesis
Raw Excerpt
“Starting from a single human-annotated template for each hand and grasp type, our pipeline tackles the complicated synthesis problem with two stages: optimize the object to fit the hand template first, and then locally refine the hand to fit the object in simulation.”