Analysis and Optimization of Energy Conversion in Bipedal Locomotion

Abstract

This thesis investigates energy-efficient dynamic locomotion for the DLR C-Runner bipedal robot using reinforcement learning. Starting from an existing walking controller, three key extensions are developed: (i) a reformulation of the locomotion problem as a partially observable Markov decision process (POMDP), incorporating a finite observation-action history into the policy inputs to enable tracking of higher forward velocities; (ii) an outer optimization of the gait timing parameters - swing ratio and period - with respect to the Cost of Transport (CoT), performed via parallel rollouts in simulation; and (iii) the integration of the motor operating region (MOR) into the training objective through a neural network approximation of the motor efficiency map. The trained running policies transfer successfully to the real robot and demonstrate a visible flight phase, a defining characteristic of running that had not been achieved on this platform before. The MOR-aware policy achieves a lower CoT than the baseline, confirming that accounting for the true motor power consumption during training leads to more efficient gaits. An energy analysis reveals that the series-elastic actuators at the ankle joints undergo compression-decompression cycles during stance, storing and releasing elastic energy during push-off. A comparison between walking and running gaits shows that running becomes more CoT-efficient above approximately 1.1 m/s on the C-Runner system. This experimentally determined walk-to-run transition point is consistent with the biomechanical principle that running is more economical at higher speeds, and confirms that the transition speed is robot-specific and cannot be derived from human reference data alone.