EC-Diffuser: Multi-Object Manipulation via Entity-Centric Behavior Generation
Authors: Carl Qi, Dan Haramati, Tal Daniel, Aviv Tamar, Amy Zhang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To study the above, we evaluate our method on 7 goal-conditioned multi-object manipulation tasks across 3 simulated environments and compare against several competitive BC baselines learning from various image representations. |
| Researcher Affiliation | Collaboration | 1 UT Austin, 2 Technion, Israel Institute of Technology, 3 Brown University, 4 Meta AI |
| Pseudocode | No | The paper describes the architecture in text and with Figure 1, but does not include any explicit pseudocode blocks labeled "Pseudocode" or "Algorithm", nor does it present structured steps formatted like code. |
| Open Source Code | Yes | We provide the project code to reproduce the experiments at https://github.com/carl-qi/EC-Diffuser. |
| Open Datasets | Yes | The datasets used are publicly available, and the work aims to improve the efficiency of robotic object manipulation tasks. |
| Dataset Splits | Yes | We subsample 3000 episodes from the real robot dataset, each padded to 45 images, and we randomly select 2700 episodes for training and 300 for validation. |
| Hardware Specification | Yes | For GPUs, we use both NVIDIA RTX A5500 (20GB) and NVIDIA A40 (40GB), though our model training requires only around 8GB of memory. |
| Software Dependencies | No | The paper mentions using specific codebases like "PINT modules from DDLP Daniel & Tamar (2024)", "Diffuser Janner et al. (2022) codebase", "VQ-Be T (Lee et al., 2024)" and "ECRL (Haramati et al., 2024)". However, it does not provide specific version numbers for these or other key software components (e.g., Python, PyTorch, CUDA). |
| Experiment Setup | Yes | Table 5: Hyper-parameters used for the EC-Diffuser model. Batch size 32 Learning rate 8e-5 Diffusion steps 5, 100 (generalization tasks) Horizon 3 Number of heads 8 Number of layers 6, 12 (generalization tasks) Hidden dimensions 256, 512 (generalization tasks) |