SO-101 Bolt Insertion with Human-in-the-Loop Corrections
Real-robot dataset for a single contact-rich task on an SO-101 arm: pick up a brass bolt from the table and insert it into the hole of a black peg. It contains 50 teleoperated demonstrations and two rounds of 50 policy rollouts in which a human operator took over whenever the policy was about to fail.
Task string: Pick up the brass bolt from the table and insert it into the hole of the black peg.
| Robot | SO-101 follower arm (6-DoF incl. gripper), SO-101 leader arm for teleoperation |
| Episodes | 150 |
| Frames | 161,787 (about 90 minutes) |
| Rate | 30 fps |
| Cameras | front (fixed, top-down view of the workspace) and wrist (gripper-mounted), 640Γ480, AV1 |
| Format | LeRobot dataset v3.0 |
Collection protocol
| Episodes | Source | Collector | Success | Human-controlled frames |
|---|---|---|---|---|
| 0β49 | Teleoperated demonstrations | human | 50 / 50 | 100% |
| 50β99 | HIL round 1 | SmolVLA fine-tuned on episodes 0β49 (smolvla_core50_r0), human takeover |
42 / 50 | 24.8% |
| 100β149 | HIL round 2 | SmolVLA trained with advantage conditioning on rounds 0β1 (smolvla_acp_r1), human takeover |
43 / 50 | 29.4% |
In the HIL rounds the policy controls the arm by default. The operator holds the leader arm and can take over at any time; while the takeover is active the recorded action comes from the leader arm, and the policy's own output is still stored in complementary_info.policy_action. Episodes were capped at 90 s. Success was labeled by the operator at the end of each episode and counts an episode as successful when the bolt ended up in the hole, with or without human help.
Object positions were varied by hand between episodes. All data was collected in one scene with one bolt and one peg.
Features
| Key | Shape | Description |
|---|---|---|
observation.state |
6 | Follower joint positions: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper |
action |
6 | Executed joint position targets, same order as the state |
observation.images.front |
480Γ640Γ3 | Front camera |
observation.images.wrist |
480Γ640Γ3 | Wrist camera |
complementary_info.policy_action |
6 | Policy output at this step (zeros in teleoperated episodes 0β49) |
complementary_info.is_intervention |
1 | 1 when the human is in control, else 0 (0 throughout episodes 0β49, which had no policy) |
complementary_info.state |
1 | Takeover state machine: 0 = policy, 1 = human takeover active, 2 = takeover being released |
complementary_info.collector_policy_id |
1 | human, smolvla_core50_r0 or smolvla_acp_r1 for the controller at this frame |
Joint values use LeRobot's calibrated normalization for SO-101 (use_degrees=False): arm joints are mapped to roughly [-100, 100] over each joint's calibrated range, and the gripper to [0, 100]. Episode-level success labels are stored in meta/episodes under episode_success (success / failure).
Usage
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("EdenYeager/so101-bolt-insert-hil")
frame = ds[0]
print(frame["action"], frame["observation.images.front"].shape)
# Teleoperated demonstrations only
demos = LeRobotDataset("EdenYeager/so101-bolt-insert-hil", episodes=list(range(50)))
Limitations
- One task, one scene, one set of objects; not meant for generalization benchmarks.
- Human takeover timing reflects one operator's judgment and is not a calibrated failure signal.
- Failure episodes (15) are kept. Filter on
episode_successif you need successful trajectories only.
License
CC BY 4.0.
- Downloads last month
- 93