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..
Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges
Authors: Tao Zhong, Jonah Buchanan, Christine Allen-Blanchette
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method across diverse hand-object combinations, demonstrating that our approach generates stable, physically grounded grasps with strong generalization. This work enables semantic grasp transfer for heterogeneous manipulators and bridges vision-based grasping with probabilistic generative modeling. Additional details at grasp2grasp.github.io. 5 Experiments 5.1 Experimental Setup Dataset. We evaluate our method on the Multi Gripper Grasp dataset [12]... |
| Researcher Affiliation | Collaboration | Tao Zhong1, Jonah Buchanan 2,3, Christine Allen-Blanchette1 1Princeton University, 2San Jose State University, 3Lockheed Martin Corporation EMAIL EMAIL |
| Pseudocode | Yes | The training and inference processes are detailed in Alg. 1, 2, and 3. Algorithm 1: Training and inference procedures. Left: VAE training. Center: Simulation-free Schrödinger Bridge training in latent space. Right: Inference by latent evolution. |
| Open Source Code | Yes | Additional details at grasp2grasp.github.io. |
| Open Datasets | Yes | Dataset. We evaluate our method on the Multi Gripper Grasp dataset [12], a large-scale benchmark containing 30.4 million grasps across 11 robotic manipulators and 345 objects. |
| Dataset Splits | Yes | For our experiments, we select 138 objects for training and 34 unseen objects for testing. |
| Hardware Specification | Yes | Our experiments were conducted on a combination of local servers and a high-performance computing cluster. The local server consists of a 24-core CPU and 2 NVIDIA A6000 GPUs, which were used for model development, ablation studies, and VAE training. For large-scale training, we utilized a compute cluster, where each cluster node is equipped with two 26-core CPUs and 8 NVIDIA L40 GPUs. |
| Software Dependencies | No | The paper mentions using "Isaac Gym [57]" and "Warp differentiable simulator [56]" and "PVCNN-based backbone [51]" and "U-Vi T model [5]" but does not specify version numbers for any of these software components. |
| Experiment Setup | Yes | We train the VAE for 18 epochs using a batch size of 256, Adam optimizer with a learning rate of 3 104, and 16 dataloader workers. The loss function includes a weighted combination of reconstruction losses with = 0.6 and a KL divergence term with a very small weight ( = 1 105). ... Table 7: Training Hyperparameters for Score and Flow Models Hyperparameter Value Total Steps 20000 Batch Size (on each GPU) 512 Learning Rate 2 104 Linear Warmup Steps 256 Gradient Clipping 1.0 EMA Decay 0.999 EMA Start Step 10000 (used in Table. 1) 0.1 |