The problems we're working on.

Most useful work in the physical world involves contact: picking things up, putting them down, fitting one thing into another. The hard parts are the things that aren't in a typical RL benchmark — friction, deformation, occlusion, and the long tail of object geometries that haven't been seen before.

We work on policies that handle these conditions, on the representations that make them learnable, and on the data and simulation pipelines that make them transferable.

Walking is solved on flat ground in a lab. It is not solved when the ground deforms, the lighting changes, the robot is carrying something asymmetric, or a sensor drops out. Locomotion in the real world is a controls problem, a perception problem, and a recovery problem all at once.

Our interest is in policies that degrade gracefully — that keep the agent upright and useful when conditions move outside the training distribution.

A useful world model lets an agent imagine what will happen before doing it. The hard problem is making models that are accurate enough to plan against, fast enough to roll out at the rate the agent needs, and general enough to be useful outside the conditions they were trained on.

We work on the representational and training questions behind that — what to predict, at what abstraction, and how to keep the model honest about its own uncertainty.

Simulation is the cheapest way to get data, and the most reliable way to get policies that don't generalize. The gap between sim and real is not one thing — it's sensor noise, contact dynamics, actuator delay, lighting, and a dozen smaller things that compound.

We build simulators, randomization schemes, and adaptation methods that close that gap deliberately rather than accidentally — and we measure transfer honestly.

A robot that picks up one object is a controller. A system that decides which object to pick up, when, in service of what goal, and how to recover when the plan fails — that's an agent. The interesting research questions sit at the seams between modules: how perception shapes planning, how plans constrain control, how failures get attributed back through the stack.

We work on the architectures that make those seams hold up under real-world conditions.

Most perception research optimises a proxy — classification accuracy, segmentation IoU, retrieval rank. Embodied agents care about something different: whether the representation supports the action that follows. A perfect segmentation that confuses a downstream grasp is worse than a noisy one that doesn't.

We study representations evaluated end-to-end on the tasks they enable, and the training regimes that produce them.

Safety in embodied AI is concrete in a way that text-only safety often isn't. A bad action has a force, a velocity, and a consequence. Safety here means evaluation methods that surface failure modes before deployment, control architectures with verifiable bounds, and an honest accounting of what the system does not know.

We treat this as a research problem in its own right — not a release-checklist item bolted onto someone else's policy.