RoboColosseum/BimanualYAM-datasets

BimanualYAM Datasets

This repository hosts teleoperated manipulation episodes collected on a bimanual YAM (bi_yam_follower) robot, recorded in LeRobot v3.0 dataset format. It contains five task subsets, each a self-contained LeRobot dataset under its own top-level folder.

Subset	Path	Task instruction	Episodes	Total frames
Dustpan	Dustpan/	Clean the table.	101	55,046
Cups	Cups/	Stack the cups.	100	39,184
Drawer	Drawer/	Put the cup into the drawer.	101	82,305
Microwave	Microwave/	Open the microwave and take out the bowl.	100	108,487
Tray	Tray/	Clean the table.	102	50,254

All five subsets share the same robot, fps, camera rig, and action/state spaces:

Robot type	bi_yam_follower (bimanual YAM arms)
FPS	30
Cameras	left, right, top (RGB, AV1-encoded video)
Action / state spaces	absolute EEF pose, delta EEF pose, joint angles

Dustpan

Episodes show a bimanual YAM setup using a dustpan-and-brush pair to clean debris off a table (one arm holds/positions the dustpan, the other sweeps debris into it), recorded under the single task prompt "Clean the table.". Episodes were filtered to remove wasted/failed demonstrations before upload (hence dustpan_filtered).

Cups

Episodes show a bimanual YAM setup stacking cups on a table, recorded under the single task prompt "Stack the cups.". 100 episodes after removing 5 episodes with bad/failed demonstrations (originally 105 recorded episodes).

Drawer

Episodes show a bimanual YAM setup placing a cup into a drawer, recorded under the single task prompt "Put the cup into the drawer.". Merged from 5 separate recording sessions; the last episode of two of those sessions was dropped for being a failed/incomplete demonstration before merging.

Microwave

Episodes show a bimanual YAM setup opening a microwave and taking a bowl out of it, recorded under the single task prompt "Open the microwave and take out the bowl.". Merged from 2 separate recording sessions.

Tray

Episodes show a bimanual YAM setup clearing a tray off a table, recorded under the single task prompt "Clean the table." (same instruction text as Dustpan, but a different object/tool - a tray rather than a dustpan). Merged from 3 separate recording sessions; episode 21 of the third session was dropped for being a failed/incomplete demonstration before merging.

Data fields

Each frame (in every subset) carries:

• Images (video, 3 channels, AV1/yuv420p, 30 fps):
  • observation.images.left — 480×640
  • observation.images.right — 480×640
  • observation.images.top — 360×640
• State:
  • observation.state_eef_absolute (16): per-arm end-effector x, y, z, qw, qx, qy, qz, gripper for left_eef and right_eef
  • observation.state_joint_angles (14): per-arm joint_0..joint_5, gripper for left/right
• Action:
  • action_eef_absolute (16): same layout as observation.state_eef_absolute
  • action_eef_delta (16): per-step delta EEF pose for each arm
  • action_joint_angles (14): same layout as observation.state_joint_angles
• Bookkeeping: timestamp, frame_index, episode_index, index, task_index

Directory layout

Each subset follows the standard LeRobot v3.0 layout, e.g. for Cups/:

Cups/
├── data/chunk-000/file-000.parquet        # low-dim obs/action/state, one row per frame
├── videos/
│   ├── observation.images.left/chunk-000/file-000.mp4
│   ├── observation.images.right/chunk-000/file-000.mp4
│   └── observation.images.top/chunk-000/file-000.mp4
└── meta/
    ├── info.json        # schema, fps, shapes, video codec info
    ├── stats.json        # per-feature dataset statistics
    ├── tasks.parquet     # task index -> text instruction mapping
    └── episodes/chunk-000/file-000.parquet  # per-episode metadata

(Some subsets have more than one video chunk/file due to size rotation, e.g. Drawer/ and Microwave/ have an extra observation.images.top file.)

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="RoboColosseum/BimanualYAM-datasets",
    root="Cups",  # or "Dustpan" / "Drawer" / "Microwave" / "Tray" - or download and point root at the local folder
)

(Requires lerobot with codebase_version = v3.0 dataset support.)

Collection

Collected via teleoperation on a bimanual YAM follower rig using this lab's lerobot fork (bi-yam recording/eval scripts), with absolute EEF, delta EEF, and joint-angle action spaces all recorded simultaneously so downstream policies can pick any action representation.