THUgewu's picture
Add TsFile (converted from LSY-lab/stack_without_ft_tact_v4)
c8e28f1 verified
|
Raw
History Blame Contribute Delete
3.32 kB
---
task_categories:
- robotics
tags:
- tsfile
- timeseries
- time-series
- LeRobot
- robotics
pretty_name: Stack (no F/T, tactile) v4 (TsFile)
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: data/stack_without_ft_tact_v4.tsfile
---
# Stack (no F/T, tactile) v4 (TsFile)
Apache TsFile version of [`LSY-lab/stack_without_ft_tact_v4`](https://huggingface.co/datasets/LSY-lab/stack_without_ft_tact_v4).
## Overview
A LeRobot teleoperation dataset for the manipulation task *"pick the lego block."*
Each episode records the synchronized robot proprioception and the commanded action
while an operator demonstrates the pick motion. The robot exposes a 7-joint arm with a
gripper; per-frame state includes the end-effector Cartesian pose, the joint positions,
the gripper opening, and the controller target pose.
- **Scale:** 51 episodes / 12,244 frames / 1 task.
- **Sampling rate:** 15 fps.
- **LeRobot codebase version:** v2.1.
## Schema (TsFile structure)
- **Time** (INT64, milliseconds) — per-frame timestamp, `round(timestamp * 1000)`, restarts at 0 for each episode (≈67 ms between frames at 15 fps).
- **episode_index** (TAG) — episode identifier (the device dimension). Query a single episode with `WHERE episode_index=0`.
- **task_index** (TAG) — task identifier; this dataset has a single task (`0`).
- **frame_index** (FIELD, INT64) — frame counter within the episode.
- **sample_index** (FIELD, INT64) — the source `index` global row counter, renamed.
- **observation_state_cartesian_0 … _5** (FIELD, FLOAT) — end-effector pose `[x, y, z, roll, pitch, yaw]`.
- **observation_state_gripper** (FIELD, FLOAT) — gripper opening.
- **observation_state_joints_0 … _6** (FIELD, FLOAT) — measured positions of the 7 arm joints.
- **observation_state_target_0 … _5** (FIELD, FLOAT) — controller target pose `[target_x, target_y, target_z, target_roll, target_pitch, target_yaw]`.
- **action_0 … action_6** (FIELD, FLOAT) — commanded action `[x, y, z, roll, pitch, yaw, gripper]`.
Each fixed-size vector column is flattened to one FLOAT field per element, keeping the
full source column name (`.` → `_`) plus the element index.
### Conversion notes
- The source carries a redundant aggregate column `observation.state` (20-dim) that is the
exact elementwise concatenation of `observation.state.cartesian` (6), `observation.state.gripper` (1),
`observation.state.joints` (7) and `observation.state.target` (6). It is **dropped** in this
TsFile version — no information is lost, because the four named sub-columns above reconstruct it exactly.
- The source `timestamp` column is dropped because it equals `Time / 1000` seconds.
- Camera video streams (`observation.images.primary`, `observation.images.wrist`) are **not**
included in this repository; see the original dataset's `videos/` directory for them:
[LSY-lab/stack_without_ft_tact_v4](https://huggingface.co/datasets/LSY-lab/stack_without_ft_tact_v4).
## Usage
Read the `.tsfile` files with the Apache TsFile Java or Python SDK.
## Source & license
- Original dataset: https://huggingface.co/datasets/LSY-lab/stack_without_ft_tact_v4
- Author / publisher: LSY-lab
- Paper: none
- License: not declared by the original dataset; please defer to the original.