File size: 2,903 Bytes
5502fb9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | ---
license: mit
task_categories:
- robotics
tags:
- robot-manipulation
- mujoco
- aloha
- world-model
- imitation-learning
pretty_name: IWS Rotate-T 90° Demonstrations (1k)
---
# IWS Rotate-T 90° Demonstrations (1k)
1,000 scripted expert demonstrations (+100 validation) of the **rotate-the-T-90°-clockwise**
task on the bimanual ALOHA push-T MuJoCo environment, in the Interactive World
Simulator (IWS) HDF5 format. Collected with
`scripts/data_collection/collect_rotate_t.py` from the
[interactive_world_sim](https://github.com/bingqingchen/interactive_world_sim) repo.
The T spawns **upright** (θ = 0) and both arms rotate it ~90° clockwise with a
closed-loop scripted policy (≤30° sub-rotations with angle feedback, so push
slippage is absorbed). Every episode runs the post-reset stabilization
(`stabilize_t`: the T falls from z = 0.07 and settles on the table **before**
anything is recorded). Only demos that actually rotated 80–105° CW while staying
flat are kept.
## Splits and blocks
| Split | Episodes | Block | Start condition |
|---|---|---|---|
| `train/` | `episode_0 … episode_199` | fixed | T pinned at (0, 0), arm at fixed reset home (`--no_settle`) |
| `train/` | `episode_200 … episode_999` | random | T at random XY in ±0.08 m, randomized arm ready pose (`settle_arms`) |
| `val/` | `episode_0 … episode_19` | fixed | same as fixed train block, disjoint seeds |
| `val/` | `episode_20 … episode_99` | random | same as random train block, disjoint seeds |
Collection seeds (`ep_seed = seed·10⁶ + trial`): train fixed 11–12, train random
21–28, val fixed 31, val random 41 — disjoint from the earlier `rotate_t` /
`rotate_t_fixed` datasets (seeds 0, 100) and from the tight-eval protocol (seed 7000).
## Episode schema (HDF5)
Identical to the IWS world-model MuJoCo dataset — drop-in for both world-model
training and BC:
```
action (T, 4) float32 bimanual EE-XY targets [Lx, Ly, Rx, Ry]
env_state (T, 7) float32 T-block pose (xyz + wxyz quat)
obs/ee_pos (T, 2, 4, 4) float32 EE poses (left, right)
obs/images/top_pov (T, 128, 128, 3) uint8 top-down RGB
obs/joint_pos (T, 14) float32 both arms' joint positions
robot_bases (T, 2, 4, 4) float32 world_T_base (left, right)
```
Episode length is variable (multiples of 60 control steps at 10 Hz — one
sub-rotation each). `videos/` inside each split holds a 128×128 mp4 preview per
episode.
## Download
```bash
python scripts/download_data_hf.py --repo jacob3333/interactive-world-sim-rotate-t-data \
--local_dir data/rotate_t_1k
# or
hf download jacob3333/interactive-world-sim-rotate-t-data --repo-type dataset \
--local-dir data/rotate_t_1k
```
Point IWS training at it with `dataset.dataset_dir=data/rotate_t_1k` (the loader
reads `train/` and `val/` subdirectories).
|