Researchers led by Wai Ka Sun have built passage-conditioned navigation, which makes a legged robot remember how it entered each stair or ledge during a 27-minute teaching session, according to a paper on arXiv. If the robot is more than 10 cm or 10° away from the taught crossing as it reaches a ledge, a gate makes it turn or back up and try again.

Under localisation noise, the share of disturbed ledge crossings within those limits rose from 24.1% to 88.6%. In contact simulation, a robot knocked off course before a staircase climbed it 180 times without a fall. It fell nine times when the stored direction was ignored. On a real wheeled-legged robot it completed all six runs over 271 m of obstacles and took third place in the 2026 Global Open-source AI Challenge. No other finalist in its task finished fully autonomously.

Source: arXiv Robotics (cs.RO)

Background

Wai Ka Sun

Robotics researcher

Robotics researcher.