metadata
license: mit
task_categories:
- robotics
tags:
- push-t
- diffusion-policy
- imitation-learning
- teleoperation
size_categories:
- n<1K
Push-T: 200 Teleoperated Demonstrations
200 demonstrations of the 2D Push-T task (push a T-shaped block onto a fixed target pose), collected by mouse teleoperation with demo_pusht.py from the Diffusion Policy codebase. Collected for a study of how Diffusion Policy performance scales with the number of demonstrations.
- Code and write-up: https://github.com/manhasv/pusht200
- Trained checkpoints (50/100/150/200 demos): https://huggingface.co/manhasv/diffusion-policy-pusht200
Contents
| Episodes | 200 (episode i starts from env seed i, seeds 0–199) |
| Total steps | 25,269 at 10 Hz |
| Episode length | 126 ± 46 steps (min 31, median 119, max 275) |
| Outcome | every episode ends in success: demo_pusht.py stops an episode automatically once the T covers > 95% of the target |
| Demonstrator | a single person, mouse teleoperation |
Format
A zarr ReplayBuffer (the format used throughout the Diffusion Policy repo). All episodes are concatenated along the first axis; meta/episode_ends gives the end index of each episode.
| Key | Shape | Type | Description |
|---|---|---|---|
data/action |
(T, 2) | float32 | target position of the agent, in pixels (0–512) |
data/state |
(T, 5) | float32 | [agent_x, agent_y, block_x, block_y, block_angle] |
data/keypoint |
(T, 9, 2) | float32 | 9 keypoints of the T block, in pixels |
data/img |
(T, 96, 96, 3) | uint8 | rendered RGB observation |
data/n_contacts |
(T, 1) | float32 | number of agent–block contacts |
meta/episode_ends |
(200,) | int64 | cumulative episode end indices |
Download
hf download manhasv/pusht-200-data pusht_200.zarr.tar.gz --repo-type dataset --local-dir .
mkdir -p data && tar xzf pusht_200.zarr.tar.gz -C data/ # -> data/pusht_200.zarr
Notes
- This dataset is for demo purpose only. Compared with the original paper's Push-T demonstrations, this dataset have similar episode lengths but more idle frames (13% vs. 3%), jerkier motion, and ~27% longer agent paths. Policies trained on this data scored lower than the authors' results at similar dataset sizes.
License
MIT, following the upstream Diffusion Policy repository.