EvolvingGrasp: Evolutionary Grasp Generation via Efficient Preference Alignment

Conference: ICCV 2025 arXiv: 2503.14329 Code: https://evolvinggrasp.github.io/ Area: Robotics Keywords: Dexterous Grasping, Preference Alignment, Consistency Models, Diffusion Models, Physical Constraints

TL;DR

This paper proposes EvolvingGrasp, which achieves efficient evolutionary generation and human preference alignment for dexterous grasp pose synthesis via Handpose-wise Preference Optimization (HPO) and a Physics-Aware Consistency Model (PCM), attaining state-of-the-art performance on four benchmark datasets with a 30× inference speedup.

Background & Motivation

The generalization capability of dexterous robotic hands in complex environments is constrained by insufficient diversity in training data. The infinite variety of real-world scenarios makes it impractical to predefine all grasping strategies. Existing approaches can be categorized as: - Optimization-based methods: optimize hand poses via force-closure criteria, but incur high computational cost. - Learning-based methods: directly regress mappings from features to grasp poses, but suffer from mode collapse. - Generative methods: e.g., DexGrasp Anything employs diffusion models, but requires hundreds of sampling steps and cannot align with human preferences.

The core challenges are: (1) existing methods cannot continuously adapt after deployment and fail to handle distributional shifts; (2) iterative sampling in diffusion models and physical constraint computation result in low efficiency; (3) no mechanism exists for aligning with human grasping preferences. Inspired by evolutionary principles—systems iteratively improve through continuous feedback from successes and failures—this paper proposes an evolutionary grasp generation framework.

Method

Overall Architecture

EvolvingGrasp consists of two core modules: 1. Handpose-wise Preference Optimization (HPO): formalizes preference alignment as posterior probability optimization, enabling the model to iteratively learn from successful and failed grasp samples. 2. Physics-Aware Consistency Model (PCM): distills a diffusion teacher model into a lightweight consistency model and integrates physical constraints to ensure physically feasible generation.

Given an object point cloud \(O \in \mathbb{R}^{N \times 3}\), the goal is to generate dexterous grasp poses with high success rates and low penetration by sampling from the posterior distribution \(P(x|O)\), where the pose parameters include joint angles \(\theta_h \in \mathbb{R}^{24}\), global translation \(T_{global} \in \mathbb{R}^3\), and global rotation \(R_{global} \in SO(3)\).

Key Designs

HPO (Handpose-wise Preference Optimization): - Introduces DPO into the dexterous grasping domain for the first time, extending it into a more flexible formulation. - Standard DPO requires paired preference data (one positive and one negative sample); HPO relaxes this constraint to allow unequal numbers of positive and negative samples. - Models preference probability via the Bradley-Terry model; the optimization objective increases the probability of successful grasps while decreasing that of failed ones. - Preference labels can be obtained via simulation evaluation (six-direction stability tests) or human-in-the-loop selection. - Employs LoRA for lightweight fine-tuning to achieve efficient preference alignment.

Physics-Aware Consistency Model (PCM): - Physics-Aware Distillation: trains a diffusion teacher model first, then distills it into a consistency student model, incorporating three categories of physical constraints into the distillation loss: - Surface Pulling Force: maintains stable contact between fingers and the object. - External Penetration Repulsion: prevents fingers from penetrating the object. - Self-Penetration Repulsion: avoids collisions between fingers. - Physics-Aware Sampling: during inference, corrects the sampling mean via gradients of physical constraints to guide trajectories toward physically feasible poses.

Loss & Training

The overall training proceeds in three stages: 1. Diffusion Pre-training: trains the teacher model with a standard noise prediction loss. 2. Physics-Aware Distillation: \(\mathcal{L}_{PAD} = \mathcal{L}_{CD} + \sum \alpha_i L_{PA_i}\), combining consistency distillation loss with physical constraints. 3. Preference Fine-tuning: applies HPO loss via LoRA lightweight fine-tuning of the entire model. - Successful and failed samples are collected from simulation; successes serve as positive examples and failures as negative examples. - Online iteration: grasp performance improves continuously as more samples are generated.

Key Experimental Results

Main Results

Evaluation on four datasets—DexGraspNet, MultiDex, RealDex, and DexGRAB:

Method	DexGraspNet Suc.6↑	MultiDex Suc.6↑	RealDex Suc.6↑	DexGRAB Suc.6↑	Time↓
UniDexGrasp	33.9	21.6	27.1	20.8	0.46s
DexGrasp Any.	53.6	72.2	34.6	56.5	32.91s
Ours w/o HPO (4-step)	63.8	75.3	51.6	55.6	1.41s
Ours (4-step)	65.2	76.8	50.6	57.7	1.41s
Ours (8-step)	65.4	80.3	64.4	60.8	2.71s
Real-time (2-step)	55.2	63.7	46.5	48.9	0.06s

Compared to the prior SOTA method DexGrasp Anything, EvolvingGrasp achieves a 30× speedup (32s → 1.41s) alongside substantially improved success rates.

Ablation Study

Module contributions validated on the MultiDex dataset (4-step):

Config	CM	PGD	PGS	HPO	Suc.6↑	Pen.↓
a	✓				60.0	14.0
b	✓	✓			64.3	12.5
e	✓	✓	✓		75.3	13.1
f	✓	✓	✓	✓	76.8	13.0

• Physics-aware distillation (PGD) improves Suc.6 from 60.0 to 64.3 (+4.3).
• Physics-aware sampling (PGS) yields the largest gain, reaching 75.3 (+11.0).
• HPO preference alignment further improves performance to 76.8.

Key Findings

1. As fine-tuning epochs increase, the Suc.6 metric improves continuously and penetration depth shows an overall downward trend.
2. Starting from a suboptimal model trained on degraded data, evolutionary fine-tuning via HPO ultimately surpasses the accuracy of the original model.
3. The real-time mode without physical guidance (2-step) requires only 0.06s, making it suitable for real-time applications.
4. Successful deployment on a real ShadowHand robot validates the evolutionary grasping capability.

Highlights & Insights

1. First application of DPO to dexterous grasping, extended into HPO that removes the strict pairing requirement, making it better suited to robotic scenarios.
2. The combination of consistency models and physical constraints is elegant: few-step generation efficiency is ensured while physical plausibility is guaranteed through dual constraints at both distillation and sampling stages.
3. Evolutionary self-improvement: the model can continuously improve after deployment using its own generated success/failure samples, without requiring additional annotations.
4. The 30× speedup represents a significant practical engineering breakthrough—reducing inference time from 32s to 1.41s and enabling real-time grasping.

Limitations & Future Work

1. Preference fine-tuning may reduce generation diversity, as alignment-oriented strategies can limit the exploration space.
2. Preference data currently originate from simulation (six-direction stability tests); preference definitions may require adjustment when transferring to complex real-world scenarios.
3. Physical constraints (e.g., penetration forces) rely on known object geometry; generalization to unknown objects remains to be validated.
4. The sensitivity of LoRA fine-tuning hyperparameters (rank, learning rate) across different scenarios is not thoroughly discussed.

Related Work & Insights

• DexGrasp Anything: a physically constrained diffusion model that is slow but produces high-quality grasps; serves as an important baseline for this work.
• Diffusion-DPO: extends DPO to multi-step MDPs for diffusion model preference alignment; the direct inspiration for HPO.
• Consistency Models (CM/sCMs): the consistency model framework enables few-step sampling; this work augments it with physical constraints.
• Insight: the combination of preference learning and physical constraints is generalizable to other embodied manipulation tasks (e.g., assembly, tool use).

Rating

Dimension	Score (1–5)
Novelty	4
Technical Depth	4
Experimental Thoroughness	4.5
Writing Quality	4
Value	4.5
Overall	4

Related Papers

• [ICCV 2025] DexVLG: Dexterous Vision-Language-Grasp Model at Scale
• [ICCV 2025] Embodied Representation Alignment with Mirror Neurons
• [NeurIPS 2025] Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges
• [ICCV 2025] PacGDC: Label-Efficient Generalizable Depth Completion with Projection Ambiguity and Consistency
• [ICCV 2025] CombatVLA: An Efficient Vision-Language-Action Model for Combat Tasks in 3D Action Role-Playing Games