Locomotion & navigation
A machine that moves through the world has to answer a question at every step: can I put weight here, and what happens next if I do. Terrain is not uniform, and the interesting cases are the ones that appear rarely.
Legged locomotion is the subsystem where learned control has most clearly beaten hand-written control, and it got there through simulation at enormous scale. But simulation only contains what someone thought to model. The failures that matter (a loose kerb edge, wet tile, gravel that gives way) are discovered in the real world, labelled by people, and fed back. The long tail is not an inconvenience here; it is the entire safety case.
What this needs from people
- Traversability judged surface by surface: gravel, wet tile, a loose kerb edge.
- Continuous-frame geometry, so the map stays coherent while the sensor is moving.
- Free space and boundaries marked with the precision a manoeuvre actually needs.
- The rare event kept, not smoothed away, because that is the case that causes the accident.
The research
Where this subsystem actually stands.
Talks & demonstrations
The people building it, in their own words.
How robots learn to hike
Marco Hutter — Learning Control for Legged Robots
Where HSV fits
Our 3D LiDAR line already produces continuous-frame labelled point clouds for autonomous perception and mapping, and our parking-slot work is navigation geometry at close tolerance: slot boundaries and inner corner points that a vehicle has to trust to move safely in a confined space.
See the delivery lineThe rest of the figure