all notes
Classical navigation stacks are a marvel of engineering: build a map, localize in it, plan a path, track the path. They work. So why spend years on the alternative — a neural policy that reads raw sensors and emits velocity directly, with no map at all?
The short answer
Because the interesting failure modes live in the seams. A SLAM-plan-track stack is a pipeline of independently-optimized modules, and the robot only ever sees the world through the abstractions each module chose to keep. A learned, end-to-end policy is forced to confront the raw observation — and that is exactly where robustness to the unexpected comes from.
In my paper on safe mobile-robot navigation, the agent’s entire world is a sparse LiDAR scan plus the relative goal. No occupancy grid. The policy is the planner. What I find compelling:
- Generalization to unseen layouts. Nothing about a specific map is baked in.
- One objective. Collision-avoidance and goal-reaching are optimized jointly, not hand-tuned across modules.
- Honest difficulty. When it fails, it fails on the actual perception problem — which is the problem worth solving.
Where I want to take it
The open questions that pull me toward a PhD are about safety and verification of these policies: can we give guarantees, or at least calibrated uncertainty, for a controller that is a black box? How far can sim-to-real go before the seams reappear? And how do we measure planning competence rigorously when the planner is learned, rather than written by hand?
More notes to come as the work continues.