Kinematic Synthesis of Planar Leg Mechanisms Through Large-Scale Dataset Generation, Geometric Filtering, and Optimization

Walking is one of many basic human motor functions, yet replicating it in robotic systems remains a complex problem. Historically, the design of walking mechanisms has relied on human intuition and iterative refining. Some well-known mechanisms, like Theo Jansen, have been invented by artists rather than engineers. In this paper, we present a novel, automated pipeline that includes dataset generation, filtering, and an optimization procedure for synthesizing 1-DOF geometrically feasible walking mechanisms. Four million mechanisms were simulated and evaluated for 25 mechanism types, for a total of 100 million mechanisms. Quantitative design criteria for walking motion were identified and applied to retain a total of 23,250 valid, stable walking mechanisms. We then apply a custom optimization process to adjust near-walking mechanisms whose joints run into the ground. A custom function is used to minimize the error related to ground intersection and step uniformity. The computational generation and optimization of walking linkages demonstrated in this work aims to systematically generate a large number of design concepts for walking mechanisms. While the focus of this work is on the synthesis of mechanisms for walking robots, the same methodology could be generalized to identify mechanisms for a wide range of applications beyond walking robots.