EDGI: Equivariant Diffusion for Planning with Embodied Agents
Authors: Johann Brehmer, Joey Bose, Pim de Haan, Taco S. Cohen
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate EDGI empirically in 3D navigation and robotic object manipulation environments. We find that EDGI greatly improves the performance in the low-data regime -matching the performance of the best non-equivariant baseline when using an order of magnitude less training data. |
| Researcher Affiliation | Collaboration | Johann Brehmer Qualcomm AI Research jbrehmer@qti.qualcomm.com Joey Bose Qualcomm AI Research Mila, Mc Gill University Pim de Haan Qualcomm AI Research Taco Cohen Qualcomm AI Research |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide explicit statements about the release of their own source code for EDGI, nor does it include a link to a code repository. It only mentions using/re-implementing the codebase of Janner et al. [2022] for baselines. |
| Open Datasets | No | To obtain expert trajectories, we train a TD3 [Fujimoto et al., 2018] agent in the implementation by Raffin et al. [2021] for 107 steps with default hyperparameters on this environment. We generate 105 trajectories for our offline dataset. |
| Dataset Splits | No | The paper mentions training on an 'offline dataset' but does not specify train/validation/test splits (e.g., percentages or sample counts) for reproduction. There are no explicit details on how the data was partitioned into these sets. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instances) used for running its experiments. |
| Software Dependencies | No | The paper mentions 'Py Bullet [Coumans and Bai, 2016 2019]' and an 'implementation by Raffin et al. [2021]' (Stable-baselines3) but does not provide specific version numbers for these software dependencies, only publication year ranges or general citations. |
| Experiment Setup | No | The paper states it 'follows the choices of Janner et al. [2022]' for hyperparameters and mentions using a 'linear noise schedule' as an alternative, but it does not explicitly list the specific hyperparameter values (e.g., learning rate, batch size, number of epochs) or detailed training configurations for EDGI within the main text or appendices. |