Reproduce “Closing the Loop on the Poppy Humanoid: Bipedal Locomotion with Linear-Quadratic Control and Learned Cost Functions” on OpenResearch

About this paper

Xulin Chen, Borui He, Ruipeng Liu, Naveed Tahir, Zhenyu Gan, Garrett E. Katz

Researchers at Syracuse University developed a closed-loop control methodology enabling reliable, unassisted bipedal locomotion on the standard Poppy Humanoid robot, which has significant hardware constraints. The system combines an LQR controller with a learned cost function, improving walking success rates from 62% to 78% in an office environment and from 30% to 42.5% in a lab setting.

Key insights

• Learning a quadratic cost function via convex optimization from imperfect open-loop hardware data can effectively differentiate successful trajectories from those leading to falls, implicitly handling credit assignment.
• Achieving LQR controller stability for this specific hardware required allowing a limited violation of the strict cost separation constraint (setting the hyperparameter epsilon < 0), preventing the cost matrices from becoming singular.
• Even with low control rates and limited sensing, a residual LQR approach combined with data-driven system identification and cost learning can enable robust bipedal locomotion on resource-constrained robotic platforms.

Summary by alphaXiv.