Why physical data is different
Physical interaction data records not just what an object looks like but how actions change the state of the world. Contact, force, timing and failure become part of the learning signal.
Types of robot data
Useful datasets include teleoperation trajectories, autonomous rollouts, demonstrations, tactile observations, videos, simulation trajectories and multimodal robot logs.
Why diversity matters
Data from multiple robots can expose a learner to different morphologies and strategies. Preserving embodiment metadata is important because the same action or observation can mean different things on different platforms.
Explore the field
Continue through the research map, model directory, robot directory and company directory.