Hardware Architecture
HumanEx Suite
A modular full-body data collection system for humanoid intelligence. From egocentric vision to whole-body locomotion, HumanEx captures the complete spectrum of human motion and perception.
Egocentric Vision
Simulates human visual perspective, capturing real attention patterns and interaction intent. Provides critical input for end-to-end models with stereo RGB and infrared cameras.
Modular Wearable
Lightweight head and hand units with integrated batteries and storage for wireless operation. Full-body exoskeleton supports extended wear. Flexibly adapts from head-only to whole-body configurations.
Multimodal Sync
Hardware-level temporal synchronization ensures timestamp alignment across head, hand, and full-body exoskeleton sensors with zero drift.
Basic Configuration
Egocentric Vision Unit
The head-mounted unit captures first-person visual data with dual stereo RGB cameras simulating binocular vision for depth perception, dual infrared cameras for precise hand tracking, and an integrated 6-axis IMU for head pose and motion trajectory recording. Its lightweight design with hot-swappable batteries enables comfortable, extended data collection sessions.
Tabletop Configuration
Head Unit + Dexterous Hand Exoskeleton
Combining first-person vision with bimanual dexterous manipulation, this configuration captures real-world task scenarios with joint-level precision tracking and force feedback. Multi-modal fusion positioning integrates IR, IMU, and hand recognition for high-accuracy hand localization. Integrated wrist-mounted batteries eliminate external power requirements.
Full-Body Exoskeleton
Built for Loco-Manipulation
A multi-sensor fusion system combining joint encoders and IMUs, purpose-built for whole-body locomotion and manipulation data collection. Physical measurement of absolute joint angles with zero cumulative drift, paired with high-frequency spatial attitude compensation.
Joint Encoders
Physical measurement of absolute angles, zero cumulative drift
IMU Sensors
Spatial attitude perception, high-frequency dynamic compensation
Forward Kinematics
Multi-modal data real-time alignment and skeletal chain computation
Full-Body Pose
High-precision, occlusion-free loco-manipulation data output
Occlusion-Free by Design
Physical exoskeleton directly measures joint angles without relying on visual features. Fundamentally immune to lighting changes and limb occlusion during complex manipulation tasks.
Unrestricted Mobility
Self-contained system design eliminates external base stations. Supports simultaneous large-range locomotion and fine manipulation for full-body coordinated data collection without spatial constraints.
Data Samples
From Collection to Deployment
Our system captures diverse manipulation and locomotion data across multiple scenarios — from tabletop assembly tasks to whole-body teleoperation of humanoid robots.
Data Validation
Four-Level Verification Pipeline
A multi-tiered validation system from hardware-level calibration to expert human review, ensuring every frame meets the highest quality standards for VLA model training. Our pipeline reduces data rejection rate to below 2%.
Hardware Calibration
- Camera intrinsic and multi-sensor extrinsic joint calibration
- Timestamp synchronization and spatial coordinate unification
- Zero-point calibration and environmental baseline confirmation before each session
Automated Screening
- Auto-rejection of severe motion blur, overexposure, or underexposure frames
- Filtering of abnormal jumps, discontinuous frames, and physically impossible motions
- Completeness checks ensuring no packet loss or truncation in multimodal streams
Algorithmic Verification
- Kinematic constraint validation of skeletal chain and joint angle consistency
- Simulation replay in physics engines (MuJoCo) for digital twin verification
- Contact detection algorithms evaluating end-effector environment interactions
Expert Review
- Semantic consistency checks verifying actions match task intent and logic
- Keyframe confirmation of critical interaction moments (grasp, release)
- High-ratio sampling quality assessment by domain experts
Through our four-level verification system, data rejection rate is reduced to < 2%, providing 100% trustworthy training data for VLA models.