Collection Methodology
Training physical AI requires data that captures the complexity of real-world interactions: contact-rich manipulation, dynamic locomotion, occlusion, lighting variance, and failure modes that don't exist in simulation.
Hardware-Synchronized Sensor Fusion
All sensor streams synchronized at the hardware level with sub-millisecond precision. Dual RGB cameras (external + egocentric), proprioception at 75-920Hz, IMU at 1000Hz, and audio streams share a unified clock for temporal alignment.
Human Verification
Every episode manually reviewed for success/failure classification by domain experts. Scene annotations include object interactions, environmental context, and failure mode categorization. Inter-annotator agreement measured and reported.
Environment Diversity
Data collected across 10+ industries and environment types: retail stores, warehouses, public spaces, industrial facilities. Varied lighting conditions, floor surfaces, human density, and object distributions capture the long-tail of deployment scenarios.
Data Specifications
Production-ready formats designed for ML training pipelines.
Sensor Modalities
Data Format
Annotation Schema
• Success/failure labels with confidence scores and failure mode taxonomy
• First-frame scene descriptions including object counts, environmental context, task intent
• Per-frame object bounding boxes and segmentation masks (available on request)
• Episode-level metadata: duration, environment type, robot morphology, task category
Training Applications
Data structured for imitation learning, reinforcement learning, and foundation model pre-training.
Imitation Learning
Success-labeled trajectories with complete state-action pairs. Human demonstrations of manipulation tasks, locomotion patterns, and multi-step procedures suitable for behavior cloning and inverse RL.
Reinforcement Learning
Deploy your models on our robot fleet for real-world training and evaluation. Collect task-specific data with your policies, enabling iterative improvement and deployment validation.
Foundation Models
Large-scale multi-modal data for pre-training generalist robot policies. Vision-language-action triplets with natural language task descriptions and environment context.
Evaluation & Benchmarking
Held-out test sets across diverse environments and morphologies. Assess sim-to-real transfer, distribution shift robustness, and long-tail capability.
Get Started
Browse public datasets or request custom data collection for your specific research needs.