Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes
Authors: Jiyao Zhang, Zhiyuan Ma, Tianhao Wu, Zeyuan Chen, Hao Dong
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments in both simulated and real-world settings validate the effectiveness of our approach, demonstrating its capability to mitigate collisions while maintaining a high grasp success rate across diverse objects and complex scenes. We conduct comprehensive simulation and real-world experiments to demonstrate the effectiveness of our method. |
| Researcher Affiliation | Academia | Jiyao Zhang1,2 * , Zhiyuan Ma1,2 * , Tianhao Wu1,2, Zeyuan Chen1,2, Hao Dong1,2 1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University 2 National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University EMAIL |
| Pseudocode | No | The paper describes the methodology in prose under sections like '3.1 Contact and Collision Aware IBS for Dexterous Grasping', '3.2 Conditional IBS Generation', and '3.3 Grasp Pose Optimization with IBS Constraints' but does not present any formal pseudocode or algorithm blocks. |
| Open Source Code | No | The code will be released after acceptance. |
| Open Datasets | Yes | We use the same object datasets and simulation environments as Dex Grasp Net2.0 [11]. The object datasets consist of 60 training objects from Grasp Net1Billion [8] and 1259 testing objects from Grasp Net1Billion and Shape Net [37]. |
| Dataset Splits | Yes | The object datasets consist of 60 training objects from Grasp Net1Billion [8] and 1259 testing objects from Grasp Net1Billion and Shape Net [37]. We sample 100 scenes from the 7600 training scenes of Dex Grasp Net2 [11] for the training of the IBS generation module. And we use the full 670 scenes from the testing set of Dex Grasp Net2 [11] for the testing. The testing scenes are categorized into three density levels: loose, random, and dense. |
| Hardware Specification | Yes | We train our model on 8 NVIDIA RTX 4090 GPUs with a batch size of 64. We assess the inference efficiency of our method on a single NVIDIA RTX 4090 GPU by averaging 50 independent runs... |
| Software Dependencies | No | The paper mentions using 'Adam W optimizer' and refers to diffusion models and UNet architectures, which imply the use of deep learning frameworks like PyTorch or TensorFlow. However, it does not specify any programming language versions, library versions, or specific software packages with their version numbers. |
| Experiment Setup | Yes | We train our model on 8 NVIDIA RTX 4090 GPUs with a batch size of 64. We use the Adam W optimizer with a learning rate of 6e 5 and trained the model for 130 epochs. The training procedure takes about 2 days. Table 8 lists specific hyperparameters: IBS Volume Size 0.2m x 0.2m x 0.2m, IBS Resolution (n) 40 x 40 x 40, Number of IBS Candidates (m) 5, Number of Grasp Poses (k) 5, Weights in Contact Energy Ed (α1, α2, α3) 80, 100, 2, Weights in Overall Energy E (λ1, λ2, λ3, λ4) 5, 1, 1000, 1, Denoising Timesteps 50. |