Capability-aware planning in physics-based climbing

Rock climbing, especially bouldering, requires both rigorous route planning and precise body control to reach the goal. Recent methods have established effective ways to control body movements in a human-like fashion by leveraging imitation learning with motion capture data. While these methods can address climbing in local areas, their ability to plan over longer horizons remains limited. To specifically address the planning problem, previous methods have built global graphs and used a forward search strategy to find the best route. However, these search-based methods focus only on wall difficulty and do not take the capability of the climber into account. In this thesis, we propose a capability-aware high-level planner that can be combined with a low-level control policy to solve the planning problem with natural climbing motions. The high-level planner accounts for both wall difficulty and the capability of the low-level control policy, and it can avoid sparse areas with no holds and also filter out difficult routes based on the low-level policy’s capabilities. The low-level policy is trained using reinforcement learning (RL) to acquire climbing skills and natural motion. We evaluate our method on a variety of walls and show that it improves route completion and planning time relative to search-only and RL-only baselines.