Automated Evolutionary Gait Tuning for Humanoid Robots Using Inverse Kinematics and Genetic Algorithms

Humanoid bipedal walking remains a challenging problem due to its inherently unstable, high-dimensional dynamics. This paper presents a hybrid framework that matches a compact screw-theory kinematic model with a multi-objective genetic algorithm (GA) for automated gait tuning. By parameterizing the foot’s half-elliptical swing (horizontal and vertical speeds) and the torso pitch angle, our GA optimizes both sagittal stride length and lateral deviation in a single fitness function. We show that the GA reliably converges—achieving plateaued performance by generation 55 and near-optimal trade-offs by generation 100—across runs of up to 500 generations. A Pearson correlation analysis confirms a strong link between fitness and forward displacement ($\rho\approx$ 0.74–0.89) and a moderate link with side‐to‐side drift ($\rho\approx$ 0.57–0.68), highlighting opportunities for refined penalty design. We demonstrate sim-to-real transfer by loading the GA’s best‐found parameters (generation 421) onto a 14 degrees of freedom (DoF) humanoid, which successfully traverses 2 m in five consecutive trials without falling. Compared to traditional Denavit–Hartenberg approaches, our screw-theory formulation reduces parameter count and simplifies Homogeneous Transformation Matrix (HTM) construction, while the GA obviates the need for manual tuning or explicit dynamic equations. This work lays the groundwork for future extensions that incorporate dynamic stability criteria, sensor feedback, and adaptive multi-objective weighting to enhance robustness and efficiency further.