Dynamically Feasible Vision-Based Foothold Adaptations for Legged Locomotion

Dynamic locomotion on rough terrain is still a challenge for legged robots. To traverse complex scenarios reliably, informations from proprioceptive and exteroceptive sensors may be required. We propose a foothold adaptation criterion that evaluates the transition feasibility between contact switches. It includes dynamics to keep into account future dynamical behaviour of the robot. The criterion is devised for dynamic locomotion based on the rigid body dynamics model (RBDM). We integrate this criterion with a vision-based foothold adaptation (VFA) strategy that considers the robot kinematics and the terrain morphology. By using the proposed dynamically-feasible foothold adaptation strategy to train a convolutional neural network (CNN) to reduce computational effort, it is possible to implement it during dynamic legged locomotion in simulation and experiments.