Quadrupeds such as Boston Dynamics’ robots and ANYmal often use LiDAR, depth cameras, or RGB cameras to inspect terrain before stepping. Those sensors add cost, weight, power draw, and calibration work. Smoke and dust can degrade LiDAR and depth measurements, while poor lighting limits conventional cameras.
Researchers from IIT, TU Delft, the University of Trento, the University of Manchester, and DLR tested a lower-cost approach on a Unitree Go1. Their paper combines terrain mapping, contact detection, and state estimation using the robot’s built-in proprioceptive inputs: an inertial measurement unit, joint encoders, and foot-force measurements. The resulting estimate feeds a model predictive controller equipped with explicit safety constraints.
Proprioception reveals how the robot moves and where its feet make contact, but it provides no advance view of untouched terrain. The robot learns the ground’s shape when its feet reach it and retains that information in a local map.
Common proprioceptive methods estimate slope from the four current contact points or filter a history of foothold positions. The first approach captures little beyond the robot’s present stance. The second can blur sharp terrain features such as platform edges. Separate terrain, contact, and body-state estimators also prevent evidence from one module from correcting another.
A failure called pseudo-contact makes that separation especially costly. A foot may touch the ground while its measured force remains below a fixed threshold, causing the controller to classify the leg as swinging. The incorrect contact state then corrupts the terrain map, body-height estimate, and support pattern used for balance.