Reproduce “Contact-Implicit Model Predictive Control: Controlling Diverse Quadruped Motions Without Pre-Planned Contact Modes or Trajectories” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gijeong Kim, Dongyun Kang, Joon-Ha Kim, Seungwoo Hong, Hae-Won Park
Researchers from KAIST propose a contact-implicit Model Predictive Control framework that enables quadruped robots to autonomously discover and execute diverse multi-contact motions in real-time, without requiring pre-planned contact sequences. The framework generated complex dynamic motions like front-leg rearing and trotting on a 45 kg quadruped robot, achieving lower tracking errors and improved robustness compared to prior methods.
Key insights
- Using a relaxed complementarity constraint *only* for computing the analytical gradient in the DDP backward pass allows the optimizer to explore and discover contact-breaking motions while the forward pass maintains strict physical realism with a hard contact model.
- Carefully designed differentiable cost terms, such as those penalizing low air time or promoting symmetric control, can implicitly guide the MPC to discover desired gait patterns (e.g., four-legged trot, pacing) without requiring explicit, periodic gait definitions.
- Multiple-shooting DDP variants like FDDP significantly enhance the robustness of contact-implicit MPC for hybrid systems by explicitly handling initial state discrepancies, thereby preventing contact mode divergence and ensuring consistent trajectory generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.