File size: 1,994 Bytes
305d5b2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
task_categories:
  - robotics
language:
  - en
tags:
  - ace-ego-0
  - robocasa
  - gr1
  - lerobot
  - camera-space
  - eef
---

# RoboCasa Camera-Space EEF (ACE-Ego-0)

LeRobot v2.0 demonstrations for the 24 GR1 RoboCasa TableTop tasks used by
[ACE-Ego-0](https://github.com/ACERobotics-VLA/ACE-Ego-0) SFT. Actions and states
are stored in the observation camera frame (end-effector position + rot6d, plus
gripper and waist).

This dataset does **not** include unused conversion caches (`experiments/`) or
legacy per-dataset index files. Training reads only the files listed below and
the shipped merged q99 statistics. The training code never recomputes norms.

## Layout

```text
<task>/
  data/chunk-*/episode_*.parquet
  videos/chunk-*/observation.images.ego_view/episode_*.mp4
  meta/info.json
  meta/modality.json
  meta/episodes.jsonl
  meta/stats_gr00t.json
  meta/stats_gr00t_delta_q99_first_state.json
norm/absolute/{norm_manifest.json,merged_stats.json}
norm/delta/{norm_manifest.json,merged_stats.json}
```

Task directory names match `configs/robocasa24/datasets.yaml` in ACE-Ego-0, for
example `gr1_unified.PnPBottleToCabinetClose_GR1ArmsAndWaistFourierHands_1000`.

## Download

```bash
huggingface-cli download ACERobotics/Robocasa_Camera_Space_Eef \
  --repo-type dataset \
  --local-dir data/robocasa24
```

Point the public recipes at this folder:

```bash
python scripts/train.py \
  --config configs/robocasa24/absolute_sft.yaml \
  --data-root data/robocasa24 \
  --norm-manifest data/robocasa24/norm/absolute/norm_manifest.json
```

Delta SFT uses `norm/delta/norm_manifest.json` and the delta recipe.

## Notes

- 24 tasks, about 5.82M frames, ego-view video at the source resolution.
- Absolute and Delta SFT share the same trajectories; only the shipped merged
  statistics and the delta transform differ.
- Per-dataset `stats_gr00t*.json` files are completeness artifacts. Runtime q99
  comes from `norm/{absolute,delta}/merged_stats.json`.