Physical AI Lab

The Lab is where the hard part of Physical AI gets done: converting a messy, real task into something a model can learn from, and proving that what works in simulation still works on the robot.

Data generation

We run teleoperation sessions to capture human demonstrations, and we build synthetic datasets — RGB, depth, segmentation, 6D pose — with domain randomisation tuned to a specific site rather than a generic distribution. The output is a dataset with known coverage, not a pile of clips.

Training

Imitation learning from demonstrations, reinforcement learning in simulation, and hybrid policies where each covers the other's weakness. We keep held-out scenarios back from the start so evaluation means something.

Sim-to-real validation

Every policy is measured on real hardware against the same metrics used in simulation — task success, cycle time, collision rate, recovery behaviour. If the numbers drop on transfer, that is a finding to fix, not a footnote.

Continuous improvement

Once a system is deployed, failures from the field come back to the Lab as new scenarios. The next version is trained against the things that actually went wrong.