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..
3D Equivariant Visuomotor Policy Learning via Spherical Projection
Authors: Boce Hu, Dian Wang, David Klee, Heng Tian, Xupeng Zhu, Haojie Huang, Robert Platt, Robin Walters
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments in both simulation and the real world that demonstrate that our method consistently outperforms strong baselines in terms of both performance and sample efficiency. |
| Researcher Affiliation | Academia | 1Northeastern University 2Stanford University |
| Pseudocode | No | The paper describes the method and architecture using prose and diagrams (Figure 2), but does not contain a formal 'Pseudocode' or 'Algorithm' section or block. |
| Open Source Code | Yes | We include the complete code for data generation and all models in the supplementary material, ensuring full reproducibility of our results. A public Git Hub repository will be provided with the final version of the paper. |
| Open Datasets | Yes | We evaluate ISP on twelve robotic manipulation tasks from the Mimic Gen benchmark [40], which is widely used in previous work on closed-loop policy learning [8, 65]. |
| Dataset Splits | No | For each task, we train three independent models with different random seeds (0, 1, and 2) for each of the 100- and 200-demonstration settings. The models are evaluated 60 times throughout training using 50 fixed rollouts per evaluation. |
| Hardware Specification | Yes | Our results are measured on a single RTX 4090 GPU. |
| Software Dependencies | No | Our model consists of an SO(3)-equivariant observation encoder followed by an SO(3)-equivariant diffusion module, both implemented using escnn [3] and e3nn [57]. |
| Experiment Setup | Yes | For the simulation experiments, we follow the hyperparameter settings from prior work [65, 5]. In detail, we use an observation window of two history steps for ISP-SO(3) and one step for ISP-SO(2). In both cases, the denoising network outputs a sequence of 16 action steps, which are used for optimization during training, while only the first 8 steps are executed during evaluation. During training, input images are randomly cropped to a resolution of 76 76, while a center crop is applied at evaluation time. We train all models using the Adam W [38] optimizer with Exponential Moving Average, and adopt the DDPM [15] framework with 100 denoising steps for both training and evaluation. |