Datasets:
README
Browse files
README.md
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
task_categories:
|
| 4 |
+
- robotics
|
| 5 |
+
tags:
|
| 6 |
+
- LeRobot
|
| 7 |
+
- robotwin
|
| 8 |
+
- sim
|
| 9 |
+
- franka
|
| 10 |
+
configs:
|
| 11 |
+
- config_name: default
|
| 12 |
+
data_files: data/*/*.parquet
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# stackblocks-stage3-v1 — RoboTwin stack-blocks-two task finetune data
|
| 16 |
+
|
| 17 |
+
Stage-3 of a 3-stage curriculum (DROID CoT pretrain → playground @grip conditioning →
|
| 18 |
+
**task finetune**). 45 human-teleop episodes of the `stack_blocks_two` task (Franka,
|
| 19 |
+
RoboTwin sim) — the source dataset of the T-series / EE-noCoT ablation campaigns.
|
| 20 |
+
LeRobot **v2.1 video/h264**, schema-identical to `Numbnut/playground-stage2-v1` and
|
| 21 |
+
`Numbnut/droid-keeps-v6`:
|
| 22 |
+
|
| 23 |
+
- `observation.images.ext` = bird_eye (**the `<loc>`/@grip grounding view**), 320×240
|
| 24 |
+
- `observation.images.wrist` = right_camera
|
| 25 |
+
- `observation.state` (8,) = joint qpos (7) + gripper open (1=open)
|
| 26 |
+
- `action` (7,) = EE delta-euler + gripper (`compute_ee_delta_actions`)
|
| 27 |
+
- task = the REAL per-episode instruction (goal-directed, e.g. "Relocate red block to
|
| 28 |
+
the middle and place green block on it.")
|
| 29 |
+
|
| 30 |
+
Conversion specifics (see `droid_provenance.json`):
|
| 31 |
+
|
| 32 |
+
- **30→15fps stride 2** — tempo parity with stage-2. NOTE: the historical T-series/EE
|
| 33 |
+
repos were native-rate; per-stored-frame deltas here are ~2× those — do not compare
|
| 34 |
+
models trained here against the historical EE-noCoT/T-series numbers directly.
|
| 35 |
+
- **Static-prefix trim 0.05 rad** (T-series-proven; mean ~14 frames): `trim` maps each
|
| 36 |
+
episode's source-frame offset. Any future CoT span table must map
|
| 37 |
+
`kf_lerobot = (kf_source - trim) // stride`, dropping kf inside the trimmed prefix.
|
| 38 |
+
- **held_out = []**: stage-3 eval is closed-loop sim (`run_benchmark.py --episodes 50
|
| 39 |
+
--seed-offset 100 --step-limit 800`), all 45 episodes train. Keep
|
| 40 |
+
`cot_exclude_held_out=False` for this repo.
|
| 41 |
+
|
| 42 |
+
No CoT span table yet: the existing `annotation_v5.json` files in the source tree are
|
| 43 |
+
an older API format (per-kf A only); v5-span regeneration (template annotator) is a
|
| 44 |
+
pending, separately-reviewed step.
|
| 45 |
+
|
| 46 |
+
Built by `policy/pi05-pure/scripts/convert_franka50_to_lerobot.py` in the RoboTwin fork.
|