Structural Optimization of Lightweight Bipedal Robot via SERL

TL;DR

SERL addresses the challenge of jointly optimizing robot structure and control in rigid bipeds by integrating genetic algorithms with reinforcement learning to iteratively search the design space and learn locomotion policies. The method uses a two-stage training with privileged information to accelerate adaptation and policy distillation for deployment. Wow Orin, a lightweight nine‑DOF biped with a fishbone Bowden ankle drive and SERL-optimized leg lengths, demonstrates superior energy efficiency (Cost of Transport $COT = P/(m g v)$) and competitive speed against Cassie and Unitree H1. The results indicate that task-driven co-design via SERL is feasible and beneficial for rapid, hardware-aware structural optimization in legged robotics.

Abstract

Designing a bipedal robot is a complex and challenging task, especially when dealing with a multitude of structural parameters. Traditional design methods often rely on human intuition and experience. However, such approaches are time-consuming, labor-intensive, lack theoretical guidance and hard to obtain optimal design results within vast design spaces, thus failing to full exploit the inherent performance potential of robots. In this context, this paper introduces the SERL (Structure Evolution Reinforcement Learning) algorithm, which combines reinforcement learning for locomotion tasks with evolution algorithms. The aim is to identify the optimal parameter combinations within a given multidimensional design space. Through the SERL algorithm, we successfully designed a bipedal robot named Wow Orin, where the optimal leg length are obtained through optimization based on body structure and motor torque. We have experimentally validated the effectiveness of the SERL algorithm, which is capable of optimizing the best structure within specified design space and task conditions. Additionally, to assess the performance gap between our designed robot and the current state-of-the-art robots, we compared Wow Orin with mainstream bipedal robots Cassie and Unitree H1. A series of experimental results demonstrate the Outstanding energy efficiency and performance of Wow Orin, further validating the feasibility of applying the SERL algorithm to practical design.

Figures (11)

• Figure 1: Overview of SERL framework. We employ a two-stage training method. Figure 2 illustrates the first stage, the SERL training process. Initially, a genetic algorithm randomly initializes a population with different leg length. These individuals are then trained using reinforcement learning. Rewards serve as the fitness metric for updating the population. The algorithm eventually converges to the individual with the highest fitness, resulting in optimized leg length.
• Figure 2: Overview of adaptation framework. Figure 3 shows the second stage of training. We use the leg length results and control policy obtained from the first stage of the SERL algorithm. The control policy is distilled to obtain a control policy that can be applied to a real robot.
• Figure 3: Robot structure design.
• Figure 4: Reward and maximum fitness of comprehensive locomotion task.
• Figure 5: Thigh and shin length of comprehensive locomotion task.