weiadam8/sim_and_real_cotraining_push_t_dataset

Sim-Sim Tee — Simulated Planar Pushing Demonstrations (carbon, large)

Simulated planar-pushing demonstrations: a pusher moves a T-shaped ("tee") slider to a target pose on a table. Each frame carries two 128×128 RGB camera views (overhead and wrist), the low-dimensional pusher state, the slider pose, and the per-episode goal pose.

This is one of the simulated datasets from Empirical Analysis of Sim-and-Real Cotraining of Diffusion Policies for Planar Pushing from Pixels (Wei et al., 2025) — project page.

The data is stored as a Zarr v2 store in the Diffusion Policy ReplayBuffer layout (a flat data/ group plus meta/episode_ends).

• 520 episodes, 112,791 frames total
• Demonstrations are human teleoperation, collected by an operator driving the pusher in simulation with a gamepad controller (not scripted or planner-generated)
• 10 Hz control rate — dt = 0.1 s between consecutive frames
• Episode length: min 36, median 211, mean 217, max 510 frames (3.6 s / 21.1 s / 21.7 s / 51.0 s); ~3.13 hours of demonstration in total
• ~894 MiB on disk (blosc-compressed)

Files

sim_sim_tee_data_carbon_large.zarr/
├── data/
│   ├── action            (112791, 2)            float64
│   ├── state             (112791, 3)            float64
│   ├── slider_state      (112791, 3)            float64
│   ├── target            (112791, 3)            float64
│   ├── overhead_camera   (112791, 128, 128, 3)  uint8
│   └── wrist_camera      (112791, 128, 128, 3)  uint8
└── meta/
    └── episode_ends      (520,)                 int64

Episodes are stored back-to-back along axis 0. meta/episode_ends holds the exclusive end index of each episode, so episode i spans [episode_ends[i-1], episode_ends[i]) with an implicit start of 0 for i = 0.

Fields

Field	Shape	Space	Meaning
action	(T, 2)	(x, y)	Commanded pusher position — an absolute world-frame target, not a delta.
state	(T, 3)	(x, y, θ)	Pusher pose. θ is identically 0 throughout this store — the pusher is rotationally symmetric, so its orientation is never populated. The column is kept for shape compatibility with the slider pose.
slider_state	(T, 3)	(x, y, θ)	Slider (T-block) pose. θ in radians, wrapped to [-π, π].
target	(T, 3)	(x, y, θ)	Goal slider pose, same space as slider_state. Constant within an episode; 360 distinct goals across the 520 episodes.
overhead_camera	(T, 128, 128, 3)	(H, W, C)	Overhead RGB, uint8, HWC layout.
wrist_camera	(T, 128, 128, 3)	(H, W, C)	Wrist RGB, uint8, HWC layout.

All poses are absolute coordinates in the simulator's world frame — positions are in meters, angles in radians. T is the time axis, shared by every array.

Consecutive frames are 0.1 s apart (10 Hz), uniformly — there are no variable-rate or dropped-frame gaps to account for when building time-indexed windows.

Value ranges over the whole dataset:

Field	min	max
action	[0.374, -0.376]	[0.798, 0.215]
state	[0.374, -0.376, 0.0]	[0.798, 0.215, 0.0]
slider_state	[0.416, -0.232, -3.139]	[0.772, 0.126, 3.140]
target	[0.580, -0.042, -0.062]	[0.595, -0.026, 0.056]

Zarr Chunking

Everything is chunked along the time axis only (trailing dims are a single chunk wide), so a chunk is a contiguous block of frames.

Array	chunk shape	number of chunks
action	(2048, 2)	56
state / slider_state / target	(1024, 3)	111 each
overhead_camera / wrist_camera	(128, 128, 128, 3)	882 each
episode_ends	(520,)	1

Usage

This dataset was collected with, and is meant to be trained and evaluated with, the code released on the project website.

The store is directly consumable by ReplayBuffer.copy_from_path and by the PlanarPushingDataset loader in that codebase, with shape_meta:

shape_meta:
  action: {shape: [2]}
  obs:
    state:            {shape: [3],           type: low_dim}
    overhead_camera:  {shape: [3, 128, 128], type: rgb}
    wrist_camera:     {shape: [3, 128, 128], type: rgb}

Note that the training pipeline moves images to CHW and scales to [0, 1]; they are stored here as HWC uint8. slider_state is extra ground-truth metadata and is not read by that loader.

Naming

The store name sim_sim_tee_data_carbon_large decomposes as:

Part	Meaning
sim_sim	Legacy name. Indicates that the data was collected in a Drake sim.
tee	The slider is a T shape ("tee").
carbon	The simulation renders under carbon arc lighting.
large	The larger version of the dataset with 520 demos.

Citation

@article{wei2025cotraining,
  title   = {Empirical Analysis of Sim-and-Real Cotraining of Diffusion Policies
             for Planar Pushing from Pixels},
  author  = {Wei, Adam and Agarwal, Abhinav and Chen, Boyuan and Bosworth, Rohan
             and Pfaff, Nicholas and Tedrake, Russ},
  journal = {arXiv preprint arXiv:2503.22634},
  year    = {2025},
  url     = {https://arxiv.org/abs/2503.22634}
}