Motion planning and safe autonomy
Research¶
How do we build robots that need less and less human guidance, and can be trusted when they act on their own? I treat this as a planning problem: combining search and sampling-based motion planning with learning, predictive control and perception, so robots find good paths quickly, keep safety margins by design, and adapt as the world around them changes.
Motion planning
Good, varied paths, found quickly
Result
How can high-level guidance make sampling-based planners faster?
Superfacets, a layer between geometry and search that points planners toward direct paths.
Result
How do we generate trajectories that are meaningfully different?
Roadmap search combined with sampling-based planning to produce distinct routes, not small variations.
Result
Can a robot learn how to navigate from demonstrations?
Inverse reinforcement learning recovers the goals behind demonstrated paths and turns them into actions.
Exploring
Can people steer a planner in plain language?
A language model translates instructions into guidance over superfacets.
Open question
When is a set of trajectories diverse in a way that is useful for the task?
Interested in this question? It could be a thesis or joint work.
Open question
Can a planner explain why it chose one path over another?
Interested in this question? It could be a thesis or joint work.
Risk-aware planning
Safety built into the plan
Result
How can safety margins be built into the planner itself?
Explicit safety zones around hazards, planned with rather than checked afterwards.
Result
How can semantic maps make urban routes safer?
Graphs that know about crossings, lanes and zones, with learned costs for weighing them.
Open question
How should risk be traded against progress when the map itself is uncertain?
Interested in this question? It could be a thesis or joint work.
Open question
How should safety zones adapt to the way people actually move around a robot?
Interested in this question? It could be a thesis or joint work.
Predictive control
Re-planning as the world moves
Exploring
How far ahead should a controller look among moving people and vehicles?
Model predictive control that adapts its horizon to how dynamic the scene is.
Open question
When is a longer horizon worth its computational cost?
Interested in this question? It could be a thesis or joint work.
Perception
From seeing to moving safely
Result
Can perception measure how safe a route is?
Traffic signs, moving objects and scene segmentation combined into a safety analysis.
Open question
How should uncertainty in what the robot sees change the plan it commits to?
Interested in this question? It could be a thesis or joint work.
Working on something related?
I'm open to collaborations, co-supervision and joint proposals.