260820_eef / README.md
learner1119's picture
260820 unchanged + EEF columns (observation.eef, action.eef) + GR00T modality.json
bd1f0a8 verified
|
Raw History Blame Contribute Delete
2.77 kB
---
task_categories:
- robotics
tags:
- LeRobot
- end-effector
- task-space
- ffw-sh5
configs:
- config_name: default
data_files: data/*/*.parquet
---
# 260820_eef
[`learner1119/260820`](https://huggingface.co/datasets/learner1119/260820) **unchanged**, plus end-effector (task space) columns for both arms and a GR00T `meta/modality.json`.
ROBOTIS FFW-SH5 in MuJoCo, single task "pick and place" (saggar onto a plate), 100 episodes, 44,136 frames, 20 fps, LeRobot v2.1.
## Features
| key | shape | content |
|---|---|---|
| `observation.state` | 16 | original, unchanged: `arm_l_joint1..7`, `gripper_l_joint1`, `arm_r_joint1..7`, `gripper_r_joint1` |
| `action` | 16 | original, unchanged: commanded joint targets, same names |
| `observation.eef` | 18 | **added**: `lpalm` pose (9) + `rpalm` pose (9), each xyz (m) + rot6d, in the `arm_base_link` frame; FK of `observation.state` |
| `action.eef` | 18 | **added**: same layout; FK of `action` (commanded pose, not the next measured state) |
| `observation.images.agentview`, `initial.obj_pos`, index columns | | original, unchanged |
- rot6d = first two **rows** of the rotation matrix, flattened (GR00T `ActionFormat.XYZ_ROT6D`).
- `arm_base_link` carries the lift joint and has the same axes as world (x front, y left, z up), so the pose does not depend on torso height (lift and head joints were not recorded).
- Grippers are not converted; use the gripper slices of `observation.state` / `action`.
## GR00T `meta/modality.json`
| group | state source | action source |
|---|---|---|
| `left_arm` / `left_gripper` / `right_arm` / `right_gripper` | `observation.state` [0:7] / [7:8] / [8:15] / [15:16] | `action`, same slices |
| `left_eef_9d` / `right_eef_9d` | `observation.eef` [0:9] / [9:18] | `action.eef`, same slices |
Video `cam_head` = `observation.images.agentview`; annotation `human.task_description` = `task_index`.
A GR00T modality config picks joint groups (joint space) or EEF groups (task space) from the same dataset.
Checked with `gr00t/data/stats.py` and a relative EEF (`XYZ_ROT6D`, horizon 50) config on `left_eef_9d` + `left_gripper`.
## How it was made
FK in the lab's `ri_motion_v5` style (`merge_mjcfs` → `MuJoCoParser` → `env.forward(q, joint_names)` → `env.get_T("lpalm"/"rpalm","site")`, expressed in `env.get_T("arm_base_link","body")` with `view_in_local`), on the robot MJCF `asset/robotis_ffw/ffw_sh5_site_added_fixed.xml` of the POSCO MuJoCo benchmark repo (notebook `tools/joint_to_eef_dataset.ipynb`).
`env.forward` clips joints to their limits; this touches only joint2 near its 0° lower limit (left: 622 state / 171 action frames, ≤ 0.06°; right: 299 state frames, ≤ 0.001°).
The original columns were checked to be identical to the source parquet files.