Twu31/so101_green_bowl

SO-ARM101 — "Put the green cube in the bowl" (2 cameras)

10 teleoperated demonstrations of a single pick-and-place task on a ~$300 SO-ARM101 arm: pick up a green cube and drop it into a bowl. Recorded with two cameras — an overhead view and a wrist (hand-eye) view.

This is a deliberately small, honest dataset: it exists to document how far 10 demos actually get you on a real arm. Spoiler: not very far on its own — see Results.

Task	Put the green cube in the bowl
Robot	SO-ARM101 (so101_follower), 6-DOF
Episodes	10
Frames	2,851 (~9.5 s each @ 30 fps)
Cameras	top (overhead) + wrist (hand-eye), both 480×640, AV1
State / action	6-dim each — shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper
Format	LeRobot v2.1

The cube and bowl are moved between episodes, so the policy cannot memorize a fixed position.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Twu31/so101_green_bowl")
sample = ds[0]
sample["observation.images.top"], sample["observation.images.wrist"]
sample["observation.state"], sample["action"]

What a policy trained on this actually achieves

A π0.5 (pi0.5) VLA fine-tuned with LoRA on exactly these 10 demos was evaluated on the real arm:

Real-arm success	~20% (with action smoothing; ~0% without a wrist camera on a harder 5-cube variant)
Teacher-forcing action error	0.54° (step 3999) / 0.76° (step 2000)
Camera ablation	top peaks early (cube localization); wrist peaks at the grasp — blanking the wrist collapses the gripper-close command

Two findings that are worth more than the dataset itself:

1. The wrist camera is what makes the grasp work. An earlier single-camera attempt on a harder variant of this task learned the demos near-perfectly offline and still never grasped anything.
2. Longer training made it worse. The 3999-step checkpoint fit the demos more tightly and became measurably less sensitive to both cameras — on the real arm it failed completely, while the less-converged 2000-step checkpoint worked. Offline fit is not the metric that matters; closed-loop generalization is.

Full pipeline, diagnostics, and write-up: twu3202/soarm101-vla.

Limitations

• 10 episodes is thin. Enough to cross from 0% to "occasionally succeeds", not enough for a reliable policy. It's published as a starting point / reproduction target, not a solved task.
• Single scene, single lighting setup, single operator, one green cube and one bowl.
• No success/failure labels in the dataset — all 10 episodes are successful demonstrations.

Related

• Twu31/so101_hand_blue_napkin — the same arm, a 101-episode handover task, which is large enough to train a working policy (SmolVLA 160k, 80%+ real-arm success).

Recording

Recorded via LeRobot teleoperation (SO-101 leader → follower):

lerobot-record --robot.type=so101_follower --robot.port=COM4 \
  "--robot.cameras={ top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30} }" \
  --teleop.type=so101_leader --teleop.port=COM3 \
  --dataset.repo_id=local/so101_green_bowl \
  "--dataset.single_task=Put the green cube in the bowl" \
  --dataset.num_episodes=10 --dataset.fps=30