rm65b-stack-v0

A bimanual teleoperation dataset for stacking blocks, collected on a dual RealMan RM65-B robot setup.

Overview

Field	Value
Robot	Dual RealMan RM65-B (left + right arm)
Tasks	1
Episodes	220
Total frames	202,667
Total duration	~112.6 min
FPS	30
Format	LeRobot v3.0

Tasks

task_index	Description	Episodes	Duration
0	Stack two cuboids on the arched block, and then stack the triangular block on top of the two cuboids	220	~112.6 min

Episode Length

Stat	Frames	Seconds
Min	619	20.6
Median	905	30.2
Mean	921	30.7
Max	1478	49.3

Observation Space

State (observation.state) — float32[28]

Bimanual state, left arm followed by right arm.

Index	Name	Description
0–2	left.tcp.x/y/z	Left TCP position (mm)
3–6	left.tcp.qw/qx/qy/qz	Left TCP orientation as quaternion (scalar-first, canonical w ≥ 0)
7–12	left.joint.0–left.joint.5	Left joint angles (degrees)
13	left.gripper	Left gripper state
14–16	right.tcp.x/y/z	Right TCP position (mm)
17–20	right.tcp.qw/qx/qy/qz	Right TCP orientation as quaternion (scalar-first, canonical w ≥ 0)
21–26	right.joint.0–right.joint.5	Right joint angles (degrees)
27	right.gripper	Right gripper state

Cameras

Key	Resolution	FPS	Codec	Description
observation.images.ego	320×240	30	h264	Third-person egocentric view (downsampled 2× from 640×480)
observation.images.wrist_left	320×240	30	h264	Left wrist-mounted camera (downsampled 2× from 640×480)
observation.images.wrist_right	320×240	30	h264	Right wrist-mounted camera (downsampled 2× from 640×480)

Action Space

Action (action) — float32[16]

Bimanual action, left arm followed by right arm.

Index	Name	Description
0–2	left.tcp.x/y/z	Left target TCP position (mm)
3–6	left.tcp.qw/qx/qy/qz	Left target TCP orientation as quaternion (scalar-first)
7	left.gripper	Left target gripper command
8–10	right.tcp.x/y/z	Right target TCP position (mm)
11–14	right.tcp.qw/qx/qy/qz	Right target TCP orientation as quaternion (scalar-first)
15	right.gripper	Right target gripper command

Dataset Structure

rm65b-stack-v0/
├── data/
│   └── chunk-000/
│       └── file-000.parquet                  # All frames (202,667 rows)
├── videos/
│   └── chunk-000/
│       ├── observation.images.ego/           # 4 × file-NNN.mp4 (packed, ≤200 MB each)
│       ├── observation.images.wrist_left/    # 4 × file-NNN.mp4 (packed, ≤200 MB each)
│       └── observation.images.wrist_right/   # 4 × file-NNN.mp4 (packed, ≤200 MB each)
└── meta/
    ├── info.json
    ├── stats.json
    ├── tasks.parquet
    └── episodes/
        └── chunk-000/
            └── file-000.parquet              # Per-episode metadata and stats

Videos were recorded one MP4 per episode at 640×480, then packed into multi-episode files and downsampled 2× via teleoperation/combine_videos.py. Per-episode timestamp ranges are recorded in meta/episodes/.../file-000.parquet under videos/<key>/from_timestamp and to_timestamp.