This Robot Can Now Climb Stairs and Find Its Way Around a Building
Humanoid robots just learned to navigate multi-floor buildings without a map.
WHAT HAPPENED: Researchers trained a humanoid robot — a Unitree G1 — to navigate multiple floors on its own without a map, completing trials with over 10 meters of vertical displacement and over 100 meters of path length.
WHY YOU CARE: A standard single-stage RL navigation pipeline struggles to scale to multi-level and multi-story terrain, held back by the difficulty of complex humanoid terrain interactions such as stairs. The team split the problem in two: first training a policy on reduced order dynamics with full 3D LiDAR observations to navigate complex, multi-story terrain, then using that navigation knowledge to kickstart a policy on full-order humanoid dynamics with a frozen locomotion policy in the loop.
THE CATCH: Out-of-distribution obstacles can still challenge the navigation policy. The researchers show that applying a Poisson safety filter to the navigation policy's output recovers safety in those cases, without dropping the navigation success rate.
- A humanoid robot walked over 100 meters and climbed more than 10 meters of vertical height indoors, with no map provided.
- Researchers used a two-step training shortcut: teach a simplified 'stick figure' version first, then transfer that skill to the full robot.
- A built-in safety filter lets the robot avoid surprise obstacles without slowing it down or hurting its success rate.
Why It Matters
Stairs are the last barrier keeping delivery and helper robots stuck on a single floor of your building.