Datasets:
SO-101 7-DOF cube suite (4:3 front camera)
These are 171 teleoperated pick-and-place demonstrations on a custom 7-DOF SO-101 arm. Each one carries a colour-specific language instruction. The dataset was recorded with LeRobot (codebase v3.0) at 30 fps.
Task
There are three coloured cubes (red, white, blue) and a black bin on a cork mat. Each episode follows one instruction:
| instruction | episodes |
|---|---|
put the red cube in the bin |
50 |
put the white cube in the bin |
50 |
put the blue cube in the bin |
71 |
Cube positions vary across episodes. Some scenes have stacked or adjacent cubes.
Robot
The arm is a modified SO-101 with 6 joints plus a gripper (7 Feetech STS3215 motors). It has an extra, limited-range wrist_pitch joint between wrist_flex and wrist_roll. It was teleoperated in joint space from a matching 7-DOF leader arm, with no IK.
action and observation.state are 7-D, in degrees (gripper in its calibrated range), in this order:
shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_pitch, wrist_roll, gripper
⚠️ The order differs from stock 6-DOF SO-100/SO-101 data:
wrist_pitchis inserted at index 4. Policies pretrained on 6-DOF SO-100 data (e.g.lerobot/smolvla_base) will see the gripper and wrist_roll slots shifted.
Cameras
| key | resolution | notes |
|---|---|---|
observation.images.front |
640×480 | Logitech Brio 100. Recording at 4:3 makes the Brio crop its 16:9 sensor, so the left and right edges of the mat are cut off. |
observation.images.wrist |
640×480 | Mounted on the gripper |
For the full-FOV 16:9 version of the same task, see NilsCode/so101_rbw_100. Feature shapes differ, so the two datasets can't be concatenated directly.
Curation
curation.json lists the episode indices a human reviewer accepted (109) or rejected (62) after watching every clip. Rejected episodes are mostly stacked or tricky scenes and jerky teleoperation. Nothing was removed from the dataset. To train on the curated subset, pass the accepted list to --dataset.episodes.
Known limitations
- One operator, one workspace, one session's lighting. Lighting changes between recording and deployment strongly affected the policies trained on this data.
- Only one instruction template; the colour word is the only difference between tasks.
- Teleoperation quality varies (see the curation list).
Associated reports
This data underlies a set of technical reports on fine-tuning VLAs (SmolVLA, π0.5) on this arm. They cover colour-grounding probes, action smoothness and a deployment field report. arXiv links will be added once posted.
License
CC BY 4.0. Author: Nilesh Sutar.
- Downloads last month
- 38