PhD Dissertation Proposal: Daniel Marew, From Optimization to Learning: Developing Controllers for Dynamic Legged Robots
Content
Speaker:
Abstract:
Legged robotics is a rapidly developing field focused on creating mobile robots that use articulated limbs, or legs, for locomotion. Unlike wheeled or tracked platforms, legged robots are uniquely capable of operating effectively in environments designed for human use because their legs allow them to overcome obstacles, traverse uneven or cluttered terrain, and maintain a small physical footprint in constrained spaces. These characteristics make them suitable for diverse domains, including manufacturing, healthcare, construction, and disaster relief.
However, fully realizing these diverse applications requires achieving adaptability, agility, efficiency, and robustness comparable to those demonstrated by humans and animals. Addressing this gap necessitates advances in both hardware and, critically, the development of robust, adaptable control frameworks.
To that end, this thesis investigates real-time optimization-based motion planning, particularly Model Predictive Control (MPC), and learning-based control methods such as Reinforcement Learning (RL), to develop robust controllers for dynamic legged robots. Completed contributions include: (1) the integration of an MPC framework with a Riemannian Motion Policy (RMP) based reactive controller for online self-collision avoidance and push recovery, validated in simulations and on a custom point foot bipedal platform and (2) a biomechanics-inspired framework combining offline motion planning and reinforcement learning to execute human-like dynamic motions, validated on a dynamic soccer kicking task on humanoid platform.
Building on these foundations, anticipated contributions target emerging hardware whose dynamics are difficult to capture in current high-throughput simulators, such as legs with actuated toes, co-actuated joints, remote actuation, and compliant, closed-chain transmissions. We propose an RL framework that retains training speed while injecting missing kinodynamic priors via action parameterizations and introducing actuator and battery models for reliable sim-to-real transfer. The framework will be validated on StaccaToe, a 6-DoF robotic leg, on jumping and continuous hopping tasks. We will then extend this framework to work with a 23-DoF full humanoid platform, PresToe, which leverages StaccaToe’s leg design, including its actuated toe joints, to study how toe actuation affects efficiency and robustness during dynamic locomotion. Together, these contributions aim to advance the adaptability and reliability of legged robots for deployment in dynamic, unstructured real-world settings.
Advisor:
Donghyun Kim