ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY
Authors: Chenrui Tie, Yue Chen, Ruihai Wu, Boxuan Dong, Zeyi Li, Chongkai Gao, Hao Dong
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We systematically evaluate ET-SEED through both simulation and real-world experiments, aiming to address the following research questions: (1) Does our method demonstrate superior spatial generalization compared to existing imitation learning approaches? (2) Can our method achieve comparable performance with fewer demonstrations? (3) Is our method applicable to real-world robotic manipulation tasks? |
| Researcher Affiliation | Academia | Chenrui Tie1,2 Yue Chen1 Ruihai Wu1 Boxuan Dong1 Zeyi Li1 Chongkai Gao2 Hao Dong1 1Peking University 2National University of Singapore |
| Pseudocode | Yes | Algorithm 1 Training phase ... Algorithm 2 Inference phase |
| Open Source Code | Yes | Videos and code are available at: https://et-seed.github.io/ |
| Open Datasets | No | For each task, we set up multiple cameras to capture full point clouds of the objects to be manipulated. ... Expert demonstrations are collected by human tele-operation. |
| Dataset Splits | No | Given 25 Demos Given 50 Demos (Figure 1) and we collect demonstrations and train our policy under the Training setting (T), subsequently testing the trained policy on both T and New Poses (NP) (Section 5.1). we only provide 20 demonstrations for each task. (Section 5.2). However, specific train/validation/test splits are not explicitly provided. |
| Hardware Specification | Yes | The tasks are performed using multiple Microsoft Azure Kinect cameras and Intel Real Sense for point cloud fusion and a Franka robotic arm for execution. |
| Software Dependencies | No | We use Segment Anything Model 2 (SAM2) (Ravi et al., 2024) to segment the object from the scene and project the segmented image with depth to point cloud. No specific version number is provided for SAM2 or any other software. |
| Experiment Setup | No | We evaluate all methods using two metrics, based on 20 evaluation rollouts, averaged over 5 random seeds. (Section 5.1, Evaluation). design a model sθ to predict the applied noise in a supervised learning fashion. More explanation about 6 can be found in Jiang et al. (2024) or appendix F. (Section 4.3). Optimize loss L = loss( ˆAk 0, A0(Ak) 1) until converged (Algorithm 1). However, specific hyperparameter values (e.g., learning rate, batch size, number of epochs) are not provided. |