StableBrowse provides the data infrastructure layer for physical AI, converting raw multimodal captures—including egocentric video, depth, IMU, and tactile streams—into model-ready datasets with synchronized timestamps and calibrated geometry. Its post-processing pipeline handles sensor alignment, hand tracking and mesh generation, camera trajectory, and action labeling, delivering auditable, training-ready episodes with provenance. The platform serves robotics and embodied AI teams that need structured physical-world data rather than scattered clips.
Funding
Funding not disclosed
Founders
Product
Problem
Physical AI systems—such as robots and embodied agents—require large volumes of real-world data showing hands using tools, objects changing state, bodies moving through space, and recovery from messy physical interactions. The internet does not contain enough of these examples, and scattered video clips lack the sensor synchronization, calibration, and structured labels needed to train reliable models.
Solution
StableBrowse provides a full data infrastructure layer that handles capture protocols, sensor synchronization, calibration, depth estimation, hand tracking and mesh generation, camera trajectory, tactile streams, temporal labels, and reviewable delivery. The platform turns raw recordings into aligned outputs with synchronized timestamps, schema-valid labels, and quality-checked artifacts. Every episode ships with synchronized timestamps, calibration, action boundaries, sensor QC, manifest files, and provenance that model teams can audit. The post-processing pipeline converts multimodal capture into structured signals—aligned sensor data, calibrated geometry, and physical action labels—that can be inspected frame by frame, making the dataset ready for direct use in model training.
Target Audience
Primary customers are robotics companies, embodied AI research labs, and model development teams that need structured, high-quality multimodal physical-world data for training manipulation, locomotion, and interaction models.
Features
- Multimodal capture supporting egocentric video, stereo depth, IMU, and tactile glove inputs in a single synchronized sensor field
- Post-processing pipeline that outputs depth maps, hand tracking, hand mesh, camera trajectory, object state, action boundaries, and sensor QC reports
- Calibrated geometry with synchronized timestamps and schema-valid labels for every episode, ensuring model-ready data consistency
- Review UI assets and manifest files that enable frame-by-frame inspection and full provenance auditing by client teams
- CLI tool (`sb inspect`) for quick verification of episode details, including domain, post-processing type, validation status, and training-readiness
- Infrastructure designed for physical manipulation domains, with temporal action labels and object-state tracking integrated across capture and delivery