← Research

Research · Area 03

Train in simulation. Survive the real world.

Simulation is where embodied AI learns; reality is where it's judged. Work in this area covers the learning side of Physical AI (reinforcement and imitation learning, domain randomization, and the emerging embodied foundation models) with a relentless focus on policies that transfer.

IPAI·TV

Watch the explainer

The idea behind this lab in ninety seconds — then come back and drive it.

▶ Watch on YouTube
The work

How the work happens.

The methods behind the research.

Policy learning

RL + imitation

Learning control from reward and from demonstration, and knowing when each is the right tool.

Sim-to-real

Crossing the gap

Domain randomization, system identification, and calibrated twins that make a policy trained in sim work on metal.

Foundation models

Generalist policies

Vision-language-action models bring broad priors to control — now expected, and not the guarantee; how far they generalize, and what certifies their action before it commits, is the open question.

Data engines

Scale that transfers

The collection, curation, and replay pipelines that turn fleet experience into training signal.

Current directions

Open problems we're pursuing.

What's being pursued now.

D1

What actually transfers

Which simulation fidelity matters for transfer, and which is wasted compute, measured, not assumed.

D2

Imitation at scale

Learning from large, messy demonstration sets without a human in the reward loop.

D3

Generalist policies for flight

Adapting vision-language-action models to the dynamics of real airframes — with a certificate that gates the commanded maneuver before it commits, not a safety envelope bolted around it.

Live lab

A generalist policy, one loop at a time.

Vision-language-action: choose an instruction and watch a policy turn pixels and words into an executed action — perception, then grounding, then a chunk of motion. The policy proposes; what makes the commit safe is a certificate on the action's own energy — agency, orthogonal to how broad the priors are.

Open the full view ↗Pick an instruction and watch vision, grounding, then an executed action chunk. Grounding shown as attribute-match; an illustrative model.

Get involved

Work in this area.

Open positions, including the Physical AI Investigator Program, are listed on Careers. Research here can also spin out into a company.

Open positions →Spin it out →