Humanoid Navigation in Uneven Terrain using Learned Estimates of Traversability

In this paper we explore discrete search-based contact space planning for humanoids using both palm and foot contact in complex unstructured environments. With a high branching factor and sparse contactable regions, it is challenging for the planner to find a contact sequence in such environments quickly. Therefore, we propose to learn a function which predicts traversability-a measure of how quickly the contact space planner can generate contact sequences to traverse a certain region. By including a learned traversability estimate into the heuristic function of the contact space planner, we can bias the planner to search the areas with more contactable regions, and thus find contact sequences more efficiently. In this paper we propose and evaluate two kinds of feature vectors for estimating traversability: Exact Contact Checking (ECC) and Approximate Contact Checking (ACC), which make different trade-offs between speed and accuracy. The experimental results show that the proposed approach using ACC outperforms both ECC and the baseline heuristic for contact space planning; ACC increases the planning success rate by 19% and reduces average planning time by 24% compared to the baseline in difficult environments with uneven terrain.