Discovering, Synthesizing, and Learning Climbing Movements

The public defense on 8th June 2020 at 12:00 will be available via remote technology. Link: https://aalto.zoom.us/j/64234063678 Zoom Quick Guide: https://www.aalto.fi/en/services/zoom-quick-guideAlthough animation is commonly captured from real humans or edited manually through keyframes and interpolation curves, such techniques are time-consuming and expensive. Because of this, a long-standing goal of computer animation research has been to synthesize movements algorithmically, e.g., through simulating the human biomechanics and framing the animation problem as optimization of joint torques over time. During recent years, such approaches have enjoyed success with movements like bipedal locomotion, but some complex movement skills have remained challenging. This dissertation focuses on synthesizing natural looking and human-like climbing movements in indoor bouldering, a popular and rapidly growing sport that recently was approved to the 2020 Tokyo Olympics together with speed and sport climbing. Indoor bouldering is a form of climbing that takes place relatively close to the ground on top of a soft landing surface, and does not require special equipment other than climbing shoes. Bouldering routes are usually short and they focus on complex climbing moves that require both strength and coordination. Planning and discovering the optimal or at least possible sequence of moves from the ground to the top hold is a challenging problem. The problem gets even more complicated when the planning should consider the body types of users such that the planned path and synthesized motions would be feasible for them. This thesis proposes a high-level path planner and low-level controller for synthesizing physically plausible and human-like movements. The high-level graph-based path planner is responsible for planning a sequence of movements to the top hold while the low-level controller synthesizes the movement details through optimizing the joint actuations of a physics simulation model of a humanoid climber. Such a low-level controller might fail to follow the planned movements; the thesis proposes ways to handle this uncertainty through low-level and high-level controller interaction. In subsequent work, the approach is developed further by employing neural networks in both supervised and reinforcement learning settings. The methods proposed in the thesis result in high-quality climbing animations without needing any reference animation or motion capture data. The work should also have applications in synthesizing other types of movements with similar characteristics, e.g., creating parkour animations based on desired footstep patterns