Learning Traversability using Contact Transition Models for Humanoid Navigation in Uneven Terrain

In humanoid robot locomotion, using palm contacts can help the robot locomote more robustly under disturbance or when traversing unstructured environments, such as disaster sites. However, adding more contacts complicates the problem by increasing the branching factor of search-based planning significantly. Furthermore, balance checking with multiple noncoplanar contacts is also expensive to compute. Using discrete search-based planners [10], [13] to compute a sequence of contacts thus becomes inefficient because the high branching factor and expensive state validity check leads to very large computation times. Our approach to overcoming this problem is to create a more informative heuristic for discrete searchbased planners that consider both foot and hand contacts. In unstructured environments with uneven terrain conventional distance-based heuristics do not capture the environment geometry, and may cause the planner to be stuck in a cul-desac due to a lack of contactable surfaces in the area. Therefore, we propose to estimate the traversability of an area in the environment as part of the planner’s heuristic. In order to be effective, our estimate of traversability needs to be computed quickly, but this is difficult because computing whether the robot can truly move through a certain area requires expensive geometric tests, collision checks, and balance checks. Instead we propose a learning framework to estimate the traversability of a region in the environment. Our results suggest that using our estimate of traversability as part of the heuristic improves the success rate of navigating through difficult environments with uneven terrain. However, the time needed to compute traversability is still significant and we seek to further reduce this time in future work.