Datasets:
T1_positionalshift
100 additional SO-101 demonstrations of T1 with a red cup at shifted positions, collected to study recovery from spatial shift in a coffee-making task. The collection covers 10 directions spaced 36° apart, with 10 episodes per direction.
This is a supplement to aailabkaist/so101_coffee_all. It contains the additional demonstrations only; the original dataset is not included or modified.
Collection and research purpose
These demonstrations were collected in response to revision item 4 for the paper's Exp5 spatial shift experiment: add demonstrations at shifted positions, retrain, and measure whether success at shifted positions recovers, showing success at the unshifted default position as a reference in the same figure.
The collection is restricted to Red–T1, the initial cup pickup task. As confirmed by the collector, the cup and blue circle were shifted together, and the remaining conditions matched the existing Exp5 T1 setup. The language instruction remains pick up the cup on the blue circle.
| Collection condition | Value |
|---|---|
| Cup | Red |
| Subtask | T1: pick up the cup on the blue circle |
| Position condition | Shifted cup and blue circle |
| Directions / angular spacing | 10 / 36° |
| Episodes per direction | 10 |
| Total episodes | 100 |
The stored metadata does not specify the displacement radius, the absolute angular origin, or a per-episode direction label. Accordingly, no episode-to-angle mapping is asserted here. episode_index alone should not be treated as a verified direction label.
Overview
| Property | Value |
|---|---|
| Episodes / action frames | 100 / 24,103 |
| Duration from action frames | 803.43 seconds (13.39 minutes) |
| Episode length | Mean 241.03 frames (8.03 s), median 239, range 201–288 |
| Episode indices | 0–99 |
| Distinct instructions | 1 |
| Robot type | so_follower (SO-101 setup) |
| Action / state dimensions | 6 each: shoulder pan, shoulder lift, elbow flex, wrist flex, wrist roll, gripper |
| Cameras | observation.images.front, observation.images.wrist |
| Video | 640×480 RGB, AV1, 30 fps, no audio |
| Sampling rate | 30 fps |
| Format | LeRobotDataset v3.0 |
| Declared split | train: episodes 0–99 |
The single instruction has task_index = 0. T1 and T4 share this instruction in the parent dataset, but this additional collection is T1 only.
Files and loading
The original LeRobot layout is preserved: frame-level actions and states in data/, dataset statistics and episode/task indices in meta/, and camera recordings in videos/.
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("aailabkaist/T1_positionalshift")
Episode indices are local to this repository. Combining these demonstrations with the parent dataset requires LeRobot-aware merging and reindexing of data, episode metadata, and video references.
Data validation note
All 100 episodes have contiguous frame indices, matching action-row counts and episode lengths, finite action/state values, and timestamps at 30 fps.
The source video intervals are longer than length / fps for episodes 0, 5, 40, 50, 80, and 90, in both camera streams. The intervals are approximately 92.23, 51.70, 57.43, 81.80, 50.50, and 10.07 seconds, respectively, while their action sequences are approximately 7.33, 6.73, 8.70, 7.90, 7.87, and 7.10 seconds. The upload preserves these source files and timestamps. Confirm which video segment corresponds to the action frames before using these episodes for training; this validation does not establish their visual/action alignment.
Intended use and interpretation
Use this supplement to test the effect of adding shifted-position Red–T1 demonstrations to the existing training data. The requested evaluation compares autonomous success before and after retraining at shifted positions, alongside success at the default position.
This repository contains demonstrations, not autonomous evaluation results. It reports no retraining outcome or success-rate improvement. Independent robot rollouts are needed to measure recovery; evaluation at directions used during collection measures recovery at those positions, while claims about unseen positions require separate held-out positions. Findings from this collection should be scoped to Red–T1.
License
Apache-2.0, matching the parent dataset.
- Downloads last month
- 116