Ropedia, a physical AI startup based in the US that builds data infrastructure for robotics and embodied AI, presented six research projects during its Super Week.

The projects cover egocentric data, spatial reasoning, world modeling, and robot control. Ropedia said EgoXperience extracts camera motion, depth, hand, and object information from first-person video.

Its dataset has about 6 million frames, or roughly 50 hours of footage. HoliWorld structures 3D scenes and tracks object interactions over time.

Ropedia also said S-Agent 2.0 improved spatial reasoning scores. Its world model’s guided action selection achieved a 55% task success rate, compared with 26% for random selection.

Ropedia-A0 reached a 70.24% success rate across 42 simulated two-handed robot tasks.

By comparison, the Ego4D research dataset spans 3,000 hours of first-person video, making EgoXperience’s dataset is smaller by duration.

Egocentric footage can raise concerns about privacy and consent because it may capture bystanders, workplaces, and homes.

🔗 Source: Ropedia

Recent Ropedia developments