perception that admits doubt
Every detection reports a confidence interval, every track a covariance. The planner sees the doubt directly.
Robotics
Teaching robots to plan with uncertainty
Embodied AI research that helps robots perceive, reason, and act with safer feedback loops.
Uncertainty is part of the perception, not part of the planner
Robots that act in the real world have to recognise when they are wrong. Our embodied work pushes uncertainty back into perception so planners can react to it, then validates everything against carefully controlled real environments before feeding the surprises back into the simulator.
conventional split
perception emits point estimates, the planner is supposed to be robust to whatever they get wrong.
our split
perception emits doubt directly — confidence intervals on detections, covariance on tracks — and the planner is wired to react to that doubt.
Where the embodied loop earns its safety
sim → real → sim
We train in simulation, validate in a carefully controlled real environment, and feed every surprise back into the simulator.
a safer feedback loop
A robot that does not know it is wrong is the dangerous one. We reward systems that stop, ask, and replan when uncertainty crosses a written threshold.
The loop tightens until the gap is small
Train in simulation, validate in a controlled real environment, feed every surprise back into the simulator. Each cycle the reality gap shrinks.
The detection table is the action table
The same scene from the hero, broken down into the controller's view: detection, confidence, decision, policy class. A low-confidence unknown does not become an executed plan.
detection confidence decision policy class urgency
d1 · box 0.92 pick execute low
d2 · human 0.78 stop & yield override high
d3 · unknown 0.41 pause · ask escalate medium
Perception that admits doubt
Robots that act in the real world have to recognise when they are wrong. Our perception stack reports uncertainty alongside every detection so planners can react.
Sim, then real, then sim
We train in simulation, validate in carefully controlled real environments, and feed the surprises back into the simulator. The loop tightens until the gap is small.
cross-cuts
Embodied uncertainty is just calibration with a deadline
The same calibration discipline that surfaces in our climate forecasts and our evaluation library lands here as a control law: when the band is wide, the robot yields.