KAPEX teleop · mug-25-dim

A LeRobot-format dataset captured through KAPEX humanoid teleoperation (Manus gloves + Vive trackers, plus optional VR head cameras).

• Repo id: adhdmi/mug-25-dim
• LeRobot codebase version: v3.0
• Episodes / Frames: 100 / 16893
• Control rate: 20 fps
• State / Action dim: 25 / 25 Task instruction (per episode): Pick up the mug with the right hand and place it on top of the cube

Schema

Feature	dtype	shape	rate
observation.state	float32	(25,)	20 fps
action	float32	(25,)	20 fps
observation.images.cam_left	video (av1)	(3, 480, 640)	20 fps
observation.images.cam_right	video (av1)	(3, 480, 640)	20 fps
observation.images.cam_wrist_left	video (av1)	(3, 480, 640)	20 fps
observation.images.cam_wrist_right	video (av1)	(3, 480, 640)	20 fps

The 59-dimensional KAPEX joint layout (when full body is recorded): 3 waist + 2 head + 7 left arm + 7 right arm + 20 left hand + 20 right hand. observation.state holds the current joint positions; action holds the target joint positions issued that tick. A learned policy that emits a 59-d vector in this layout is plug-compatible with BaseEnv._apply_action in the Kapex Benchmark repository.

Cameras

All camera streams are stored as videos at 20 fps, channel-first (3, H, W) RGB layout (decoded by LeRobot's VideoFrame reader at load time).

• observation.images.cam_left — head left eye (mounted on HL2)
• observation.images.cam_right — head right eye (mounted on HL2)
• observation.images.cam_wrist_left — left wrist camera (mounted on LA7)
• observation.images.cam_wrist_right — right wrist camera (mounted on RA7)

Loading the dataset

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("adhdmi/mug-25-dim")          # downloads on first call
print(len(ds), ds.meta.fps, list(ds.meta.features))

sample = ds[0]
print(sample["observation.state"].shape)  # (25,)
print(sample["action"].shape)             # (25,)

Training

LeRobot 0.5.x ships training entry points as console scripts. Replace act with diffusion, tdmpc, vqbet, ... for other policy families.

# Action-Chunking Transformer (ACT)
lerobot-train \
    --policy.type=act \
    --dataset.repo_id=adhdmi/mug-25-dim \
    --output_dir=./outputs/act_mug-25-dim

# Diffusion Policy
lerobot-train \
    --policy.type=diffusion \
    --dataset.repo_id=adhdmi/mug-25-dim \
    --output_dir=./outputs/diffusion_mug-25-dim

Visualization

# Interactive Rerun-based viewer
lerobot-dataset-viz \
    --repo-id adhdmi/mug-25-dim \
    --episode-index 0

If lerobot-dataset-viz is not on your $PATH, run python -m lerobot.scripts.visualize_dataset --repo-id adhdmi/mug-25-dim --episode-index 0 on older releases.

Authoring info

• Recorded with scripts/tabletop_teleop/run_teleop.py.
• Converted to LeRobot format with scripts/tabletop_teleop/lerobot_pipeline/convert_record_data.py.
• Pushed to the Hub with scripts/tabletop_teleop/lerobot_pipeline/export_and_upload.py.
• Maintainer: adhdmi (Hugging Face Hub).