Funding
Funding not disclosed
Founders
Product
Problem
Training robots for dexterous manipulation requires large amounts of high‑quality human sensorimotor data, but existing tools for recording, standardizing, and delivering such data are fragmented and labor‑intensive. This limits the speed at which embodied AI systems can learn complex tasks from human demonstrations.
Solution
FPV Labs provides a full‑stack infrastructure that captures human motion and force data, processes it into robot‑ready representations, and transfers it to robot learning pipelines. The platform standardizes sensor setups, synchronizes multimodal streams, and applies automated labeling and normalization to produce consistent datasets. Processed data can be accessed through APIs or exported to common robotics frameworks, enabling developers to train embodied AI models more efficiently. By abstracting the data pipeline, FPV Labs reduces the engineering overhead for robot researchers and accelerates the deployment of robots that can imitate human manipulation.
Target Audience
Primary customers are robotics research labs, robot‑software developers, and companies building manipulation‑focused embodied AI systems that need scalable human demonstration data.
Features
- Modular hardware kits for synchronized capture of kinematics, tactile, and force signals during human demonstrations
- Automated preprocessing pipeline that cleans, aligns, and normalizes multimodal data into a unified robot‑compatible format
- Standardized representation schemas (e.g., trajectory, contact maps) that facilitate cross‑project dataset reuse
- Cloud‑based data storage with versioned datasets and metadata searchable via RESTful APIs
- Integration libraries for popular robot learning frameworks (ROS, PyTorch, TensorFlow) to stream processed data directly into training workflows
- Built‑in tools for annotating task intents and segmenting demonstrations into reusable skill primitives