File size: 2,220 Bytes
eef1706 | 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 | ---
license: mit
task_categories: [robotics]
tags: [LeRobot, vine2, delta-actions]
---
# VINE2_sim_420_dee
**Delta-EE action relabeling** of [`EndeavoringYoon/VINE2_sim_420`](https://huggingface.co/datasets/EndeavoringYoon/VINE2_sim_420).
Everything except the `action` column (videos, `observation.state`, `intervention`,
episode metadata, `meta/episode_labels.csv`, `meta/train_eval_split.json`) is copied verbatim.
## What changed
The original `action` is an 8-dim **absolute joint command** (7 arm joints + gripper).
Here it is replaced by a 7-dim **task-space delta**, following the action convention of
"Why Does Action Chunking Improve Behavioral Cloning Performance?" (Lazzati et al., 2026,
arXiv:2608.02547, Appendix A.4/A.5):
```
action[t] = [ dx, dy, dz, drx, dry, drz, gripper ]
(dx,dy,dz) = p_cmd - p_cur # translation, world frame [m]
(drx,dry,drz) = axis_angle( R_cmd @ R_cur^T ) # rotation delta, world frame [rad]
gripper = original action[7], unchanged
```
where `(p_cmd, R_cmd) = FK(commanded joints a[:7])` and `(p_cur, R_cur) = FK(measured
state s[:7])`, both evaluated at the **gripper pad centre** of the real2sim-calibrated
MuJoCo scene (`VINE2_data_collection/asset/scene.xml`, lift/head joints and base pose
fixed to the calibrated task values).
Key property: the delta is referenced to the state **at execution time** — not to the
chunk-start state (the GR00T-style convention the paper explicitly avoids). At deployment,
the absolute command is recovered as `IK( FK(current state) (+) predicted delta )`.
`meta/stats.json` and the per-episode stats in `meta/episodes/` were recomputed for the
new action; `info.json` declares the new shape `(7,)` and names.
## Why
With absolute joint actions, corr(action, state) is about 0.99 — the answer is printed in the
observation, which structurally erases non-Markov demonstrator-intent signal. After this
relabeling corr drops to about 0.42 and an ACT (chunk 50) trained on the sim variant shows a
genuine delayed-prediction advantage (best delay 2 steps at 10 Hz, -9.7 percent val error),
which the absolute version does not.
Generated by `make_delta_ee_dataset.py` (VINE2 workspace), 2026-08-27.
|