PhD Dissertation Proposal: Daniel Marew, From Optimization to Learning: Developing Controllers for Dynamic Legged Robots
Content
Speaker:
Abstract:
At its core, robotics aims to build machines that translate perception and computation into meaningful actions in the physical world, alleviating hazardous, repetitive, and physically demanding work while amplifying human capability. Legged robotics advances this mission by providing mobility aligned with infrastructure shaped by human anthropometry. Unlike wheeled or tracked platforms, legged robots are drop-in compatible with a wide range of existing settings because they can negotiate diverse terrain types, from level floors to stairs and obstacles, and maneuver through cluttered, tight corridors as well as rough, uneven terrain. These characteristics make them suitable for diverse domains, including manufacturing, healthcare, construction, and disaster relief. However, despite their promise, current legged robots remain limited in adaptability, agility, efficiency, and robustness compared to humans and animals. Bridging this gap requires not only advanced hardware but also control frameworks capable of enabling dynamic and versatile behaviors across a wide range of tasks and conditions.
This thesis investigates motion planning and learning based control methods with the overarching goal of developing robust controllers for dynamic legged robots and exploring ways to integrate these approaches. The completed contributions include: (1) a reduced-order Model Predictive Control (MPC) framework augmented with Riemannian Motion Policy (RMP) based reactive control for online self-collision avoidance and recovery maneuvers, validated through both simulations and a custom- built point-foot biped platform; and (2) a biomechanics-inspired framework that combines offline motion planning and imitation learning to generate human-like, kinodynamically feasible trajectories, demonstrated in dynamic soccer kicking tasks and shown to improve reinforcement learning training efficiency.
Building on these foundations, the anticipated contributions target emerging hardware: legs with actuated toes, biarticular/co-actuated joints, distal-mass-reducing remote actuation, and compliant, closed-chain transmissions whose dynamics are difficult to capture in high-throughput parallelized open-chain simulators. We pro- pose an RL framework that retains training speed while injecting missing kinodynamic priors via action parameterization, actuator and battery electromechanical models, and impact-mitigation strategies for reliable sim-to-real transfer. The frame- work will be validated on Staccatoe, a 6-DoF anthropomorphic robotic leg with an actuated toe and belt-driven closed chains, aiming for continuous hopping on hardware; and extended to Prestoe, a full humanoid platform, to quantify how toe actuation affects efficiency, robustness, and impact handling during dynamic locomotion. Together, these contributions aim to advance the adaptability and reliability of legged robots, enabling their deployment in increasingly dynamic and unstructured real-world settings.
Advisor:
Donghyun Kim