zero-runtime/so101_pick_place_bowl

SO-101 pick & place (orange cube → white bowl)

105 teleoperated demonstrations of an SO-101 arm that picks up an orange cube and places it in a white bowl, recorded with three cameras and the arm's joint state.

Every episode was driven by a person over the network with ZeroRuntime, recorded in the cloud, and exported to LeRobot v3.0 from the ZeroRuntime dashboard. No conversion scripts.

Fine-tuned policy: zero-runtime/smolvla_so101_pick_place_bowl

Robot	SO-101 follower (so101_follower), single arm, 6-DoF
Task text	pick up the orange cube and place it in the white bowl
Episodes	105
Frames	21,406 (about 36 min of demonstrations, ~20 s per episode)
FPS	10
Cameras	3: top, wrist, side, 640×480, AV1
Format	LeRobot v3.0 (tagged v3.0)
Recorded with	ZeroRuntime teleoperation, human leader, cloud recording
License	Apache-2.0

How it was recorded

ZeroRuntime connects a robot and its operator through an online room. The arm runs a follower process that streams joint state and camera frames, and applies every incoming command through speed clamps, a watchdog and a deadman. The operator drives a leader arm from anywhere. With cloud recording on, each time the operator takes control becomes one episode, labelled with the task, and every action is stored alongside the frames that led to it.

 operator + leader arm      ZeroRuntime room       follower.py ── SO-101 + 3 cameras
 (anywhere)          ─────── cloud recording ─────  clamps · watchdog · deadman
                                    │
                         dashboard → LeRobot v3.0

From the dashboard, the same recording also exports to RLDS, Zarr, robomimic and HDF5.

Data

State and action: observation.state and action, float32[6], SO-101 joint order:

shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos

The joint numbers are not in degrees. Each joint goes from −100 to 100, and the gripper goes from 0 (closed) to 100 (open). A value can go a little past 100, which is normal. action is where the operator told the arm to go; observation.state is where the arm actually was.

Cameras:

Key	View	Resolution
observation.images.top	top-down	640×480
observation.images.wrist	gripper / wrist	640×480
observation.images.side	side	640×480

The videos are AV1, yuv420p, 10 fps.

Indices: timestamp, frame_index, episode_index, index, task_index.

Layout:

meta/      info.json · stats.json · tasks.parquet · episodes/
data/      chunk-000/file-000.parquet
videos/    observation.images.{top,wrist,side}/chunk-000/file-000.mp4

Load it

pip install "lerobot[dataset]==0.6.1"

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("zero-runtime/so101_pick_place_bowl")
print(ds.num_episodes, ds.fps, ds.meta.camera_keys)   # 105 10 [... side, top, wrist]
frame = ds[0]   # dict of tensors: state, action, three images, indices

Browse it in the LeRobot dataset visualizer, or locally:

lerobot-dataset-viz --repo-id zero-runtime/so101_pick_place_bowl --episode-index 0

Train on it

This dataset teaches one task: picking up the orange cube and placing it in the white bowl. A policy trained on it works best on a similar SO-101 setup, with the same objects, camera views and lighting.

Reproduce the policy. These are the settings behind zero-runtime/smolvla_so101_pick_place_bowl. SmolVLA's base model expects cameras named camera1 to camera3, so they are renamed:

lerobot-train \
  --policy.path=lerobot/smolvla_base \
  --policy.train_expert_only=true \
  --policy.freeze_vision_encoder=true \
  --dataset.repo_id=zero-runtime/so101_pick_place_bowl \
  --dataset.image_transforms.enable=true \
  --rename_map='{"observation.images.top": "observation.images.camera1", "observation.images.wrist": "observation.images.camera2", "observation.images.side": "observation.images.camera3"}' \
  --batch_size=16 --steps=20000 \
  --output_dir=outputs/smolvla_so101_pick_place_bowl

Keep the camera order consistent between training and inference. Swapping two cameras raises no error; the arm simply behaves badly.

Teaching a new task needs new demonstrations of that task. Record them with ZeroRuntime (below), then fine-tune on them, from lerobot/smolvla_base or from the policy above.

Record your own with ZeroRuntime

The follower is your LeRobot robot, unmodified, wrapped by the ZeroRuntime adapter:

from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
from zeroruntime.teleops import FailsafeAction, SafetyConfig
from zeroruntime.teleops.adapters.lerobot import SO101FollowerAdapter

arm = SO101Follower(SO101FollowerConfig(id="follower", port=PORT, cameras=CAMERAS))
follower = SO101FollowerAdapter(
    MEETING_ID, robot=arm, token=TOKEN,
    safety=SafetyConfig(slew=SLEW, watchdog_timeout_s=0.5, on_starvation=FailsafeAction.HOLD),
    cloud_recording=True,
)
follower.run(hz=30)

Full scripts are in the examples repo, and the setup guide is in the docs.

Run a policy with ZeroRuntime

Once a policy is trained, ZeroRuntime runs it on the real arm over the network. The arm keeps running the same follower used for recording. The model runs as a leader on any machine with a GPU, such as a cloud GPU, a notebook, a Jetson or a laptop. The follower treats it like a human operator, so every command still passes through its speed clamps, watchdog and deadman.

 SO-101 + 3 cameras ── follower.py ──── ZeroRuntime room ──── smolvla_leader.py ── SmolVLA
 (robot)               clamps · watchdog · deadman              (any GPU, anywhere)

1. On the robot, start the follower, with the cameras labelled top, wrist and side:

python follower.py

2. On the GPU machine, check the policy first (no robot needed), then let it drive:

pip install "lerobot[smolvla]==0.6.1" "zeroruntime-teleops[lerobot]" python-dotenv opencv-python-headless

python smolvla_leader.py \
  --policy zero-runtime/smolvla_so101_pick_place_bowl \
  --dataset zero-runtime/so101_pick_place_bowl \
  --task "pick up the orange cube and place it in the white bowl" \
  --camera-map camera1=top,camera2=wrist,camera3=side \
  --check

# then the same command without --check (add --num-steps 5 on edge devices)

Press Ctrl-C once to stop. The leader lets go of control and the arm holds its pose. Full scripts are in the examples repo.

Citation

@misc{zeroruntime2026so101pickplacebowl,
  title  = {SO-101 pick and place (orange cube to white bowl), teleoperated with ZeroRuntime},
  author = {ZeroRuntime},
  year   = {2026},
  url    = {https://huggingface.co/datasets/zero-runtime/so101_pick_place_bowl}
}