Robotics studio lands 50K hours for VLA training

Overview

A Visual-Language-Action model only learns dexterous manipulation if the data behind it shows hands doing what hands actually do.

A Robotics Studio building VLA models needed a high-diversity, task-specific dataset of household dexterous manipulation ” captured on AR prototype devices in real homes, not in a lab.

Firstsource ran the program across moderated and unmoderated settings, delivering 50,000+ hours of accepted egocentric data across diverse American households and 4+ household types.

This was Intelligence that Operates: ego-exo capture running on a real-world cadence, feeding VLA model training with the diversity a controlled environment couldn't replicate.

Challenges

• Lab-only manipulation data doesn't generalize. A VLA model trained on lab capture overfits to lighting, surfaces, and object positions that real kitchens don't share. Real household environments are the training signal.
• AR prototype devices come with their own constraints. Continuous-recording windows, charge cycles, and on-device processing limits dictate what the program can deliver ” and how fast. Capture has to be designed around the device, not against it.
• Diversity at this scale isn't a sampling problem. Living rooms, kitchens, bedrooms, and bathrooms each shape manipulation differently. Coverage across 4+ household types requires intentional environment selection ” not opportunistic recording.

How We Made It Happen

We ran the program across moderated and unmoderated settings to give the VLA model the diversity it needed.

• Egocentric capture across real American households. AR-prototype-device recordings in living rooms, kitchens, bedrooms, and bathrooms ” single-family home setups representing diverse households.
• Physical AI delivered as a model-ready dataset. 50,000+ hours of accepted egocentric data, structured for direct ingestion into VLA training.
• One pipeline across moderated and in-the-wild collection. The same quality discipline applied whether capture was moderated or unmoderated ” no two-tier data inside the dataset.

Conclusion

A VLA model only generalizes when its training data sees the real world. Firstsource ran egocentric capture across diverse American households at production scale ” turning embodied data collection into Intelligence that Operates.