Datasets/Humanoid Dual-Arm Dataset

Humanoid Dual-Arm Dataset

600 hours of humanoid dual-arm manipulation across diverse household and retail scenarios. Features depth sensing, force feedback, and 4-camera coverage for complex object interactions.

RGBDepthForcecommercialFeatured
Dataset Size
300GB
Episodes3,600
Total Frames216,000
Cameras4
Data Size300GB
FormatLeRobot, HDF5, MCAP

What's Included

3,600 episodes of humanoid manipulation tasks
216,000 synchronized RGB + Depth frames
4-camera synchronized observations
Force/torque sensor data (100Hz)
Dual-arm trajectories with humanoid kinematics
End-effector 6D poses for both arms
Task labels and success indicators
Multi-format data loading utilities

Use Cases

Humanoid Robot Learning
Train policies for anthropomorphic robots
Multi-Modal Fusion
Combine RGB, depth, and force signals
Dual-Arm Coordination
Learn bimanual manipulation strategies
Contact-Rich Manipulation
Leverage force feedback for precise control
Retail Automation
Train models for shelf stocking and object arrangement
Imitation Learning
Learn from human-like manipulation demonstrations

Why trust this dataset?

Provenance, QA, and compliance details for the dataset above — what was captured, by whom, on what hardware, under what license, and what we explicitly do not claim.

Collection hardware
4-camera rig at 20 Hz, 640x480 per camera (RGB + Depth). 12-DOF humanoid-dual-arm platform with 6-D end-effector pose tracking.
Operators
Trained SignIQ teleop operators (multi-operator capture per dataset). Sub-20ms teleop latency with haptic feedback.
Environment
3 scenarios: home, retail, supermarket.
Annotation policy
Operator-written task instructions, hindsight VLM captions, reviewed subtask spans. Outcome (success/failure) labeled per episode.
QA checks
12-check automated pipeline before delivery: schema validity, timestamp monotonicity, sync drift (<5 ms), no NaN/Inf, joint-limit compliance, calibration provenance, language presence, outcome label, format export, and more.
Sensor coverage
RGB + depth + force/torque; JPEG for RGB, PNG for depth.
License
commercial delivery. Custom commercial terms available for redistribution, fine-tuning, and on-prem caching.
Privacy & consent
Operators are SignIQ staff under explicit recording consent. Faces in any incidental human capture are blurred; PII is reviewed before release.
Format validation
LeRobot, HDF5, MCAP. Each export passes a per-format validator (LeRobot info.json, GR00T modality.json, HDF5 schema, MCAP channel coverage) before bundling.
Known limitations
Single facility / lighting profile per dataset. No depth or tactile streams unless explicitly listed. Episode count and scenario coverage are advertised, not extrapolated &mdash; if you need broader coverage, ask about a custom run.