README.md

---
license: mit
task_categories:
  - robotics
tags:
  - behavior-cloning
  - diffusion-policy
  - pusht
  - bimodal
  - mode-editing
size_categories:
  - n<1K
---

# Push-T Controlled Bimodal Teleop Dataset

100 human-teleoperated Push-T demonstrations (50 LEFT-wrap + 50 RIGHT-wrap), collected
under a **controlled fixed initial state** so the two modes have kinematically equivalent
costs. Intended for downstream bimodal behavior-cloning and mode-editing experiments.

All demos reach `coverage >= 0.95` shapely-IoU success threshold (env-defined). Demos
were collected at 10 Hz mouse-driven teleop on the 2D pygame Push-T env from the
Diffusion Policy paper (Chi et al., RSS 2023).

## What "LEFT" vs "RIGHT" means

T starts at lower-left `(185, 327, π/4)` and must translate up-right to the green goal
silhouette at `(256, 256, π/4)`. The agent (blue circle) spawns upper-right of the T
at `(327, 185) ± 15 px`. Two equal-cost wrap paths exist:

- **LEFT** — agent wraps the **upper-left** side of the T (counter-clockwise about T's center)
- **RIGHT** — agent wraps the **lower-right** side of the T (clockwise about T's center)

Modes alternate by seed parity: even seed → LEFT, odd seed → RIGHT.

![Trajectory overlay](pusht_controlled_mixed_paths.png)

## Files

| File | Purpose |
|------|---------|
| `pusht_controlled_mixed.zarr/` | The dataset (zarr v2 directory store, 100 episodes / 8060 steps) |
| `pusht_controlled_mixed_paths.png` | All 100 agent paths, split by mode (visual sanity check) |
| `demo_pusht.py` | The patched collection script used to generate this dataset |

## Zarr layout

```
pusht_controlled_mixed.zarr/
├── data/                                  # 8060 timesteps total, 10 Hz
│   ├── state        (8060, 5)  float32    # [agent_x, agent_y, T_x, T_y, T_angle]
│   ├── action       (8060, 2)  float32    # teleop mouse-target action
│   ├── img          (8060, 96, 96, 3) uint8   # 96×96 RGB env render
│   ├── keypoint     (8060, 9, 2)  float32     # 9 T-block keypoints
│   └── n_contacts   (8060, 1)  float32
└── meta/                                  # 100 episodes
    ├── episode_ends       (100,) int64    # cumulative end-indices (DP standard)
    ├── episode_modes      (100,) int8     # 0 = LEFT, 1 = RIGHT  ← extra
    └── episode_coverages  (100,) float32  # final IoU coverage    ← extra
```

The `data/*` arrays and `meta/episode_ends` follow the **same layout** as the canonical
`pusht_cchi_v7_replay.zarr` from the DP repo. The two extra `meta/*` arrays are
additions for mode-editing work; the original DP training code ignores them.

## Loading

```python
import zarr, numpy as np
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id='haohw/pusht-controlled-mixed',
    repo_type='dataset',
)
r = zarr.open(f'{local_dir}/pusht_controlled_mixed.zarr', mode='r')

ends   = np.asarray(r['meta/episode_ends'])
modes  = np.asarray(r['meta/episode_modes'])    # 0=LEFT, 1=RIGHT
covs   = np.asarray(r['meta/episode_coverages'])
state  = np.asarray(r['data/state'])
action = np.asarray(r['data/action'])

# Per-episode iteration
starts = np.concatenate([[0], ends[:-1]])
for s, e, m in zip(starts, ends, modes):
    ep_state  = state[s:e]
    ep_action = action[s:e]
    mode_name = 'LEFT' if m == 0 else 'RIGHT'
    ...
```

For fast pulls on a cluster, set `HF_HUB_ENABLE_HF_TRANSFER=1` and install `hf_transfer`.

## Splitting into per-mode zarrs

If downstream code expects two single-mode zarr files
(`pusht_controlled_left.zarr` + `pusht_controlled_right.zarr`), the split is just a
filter on `meta/episode_modes`. The original DP collection guide's two-file format
can be reconstructed from this single mixed file at training time.

## Collection spec

**Env**: `PushTKeypointsEnv` from `diffusion_policy.env.pusht.pusht_keypoints_env`,
render_size=96, control_hz=10.

**Per-episode init** (overrides env's random init via `env.reset_to_state`):

| Object | Position | Angle |
|--------|----------|-------|
| T block | fixed `(185, 327)` | fixed `π/4` |
| Goal   | fixed `(256, 256)` (env default) | fixed `π/4` |
| Agent  | `(327, 185) + uniform(-15, +15)` (seeded by episode index) | — |

**Success**: shapely-IoU between current T polygon and goal T polygon `>= 0.95`.
DP paper reports continuous max-coverage per episode (clipped at 0.95), not binary
success rate — so 95% is the env's done flag, not the eval metric.

## Episode-length stats

Mean 80.6 steps (≈8.1 s @ 10 Hz), median 78, min 60, max 141.

## Reproducing

See `demo_pusht.py`. Key changes from the stock DP script:

1. Controlled init via `env.reset_to_state` (T fixed, agent perturbed ±15 px)
2. Mode auto-assigned by seed parity, displayed in-window (red/blue overlay + IoU bar)
3. Extra `meta/episode_modes` and `meta/episode_coverages` saved per episode
4. `S` key to save partial demos before 95%, `C` key to cancel and restart same seed

## License

MIT. Re-uses geometry / env from
[real-stanford/diffusion_policy](https://github.com/real-stanford/diffusion_policy).