Walking over disconnected footholds—like stepping stones or scattered debris—requires a humanoid robot to decide both where to land and when to shift its weight. A new paper on arXiv presents a perceptive model predictive planner that tackles these two problems together.

The planner combines discrete region selection with foot placement and step timing. Rather than assuming a fixed walking rhythm, it adapts step timing to the robot's current motion, allowing it to pick landing regions that fit the robot's actual dynamics.

Because the approach is perceptive, it can use sensed terrain information to plan over disconnected footholds. The paper is available on arXiv under identifier 2603.07400.