At the Crossroads of Control:
Model-based VS Learning based Control for Humanoids
In recent years, humanoid locomotion control has rapidly advanced through diverse paradigms— ranging from classical inverse dynamics to whole-body MPC, hybrid model-based-learning architectures, and end-to-end reinforcement learning (RL). Each offers distinct strengths in interpretability, scalability, and real-world adaptability. This workshop will bring together researchers with hands-on experience deploying these approaches on physical humanoids to examine which control strategies hold the most promise for future applications. Key questions include: Is whole-body MPC still the most robust option for structured environments? Can hybrid approaches truly combine the best of both worlds? Are end-to-end learned policies ready for full-body, contact-rich locomotion?
Beyond benchmark metrics, we will highlight failure cases and deployment bottlenecks to evaluate not just algorithmic performance but also engineering feasibility. The goal is to explore whether the divide between model-based and learning-based control is real or fading—and what future control pipelines might require in terms of safety, adaptability, and maintainability.
Topics of interest for this workshop are:
Model-based control methods and whole-body MPC for legged/humanoid robots
Reinforcement learning for locomotion and contact-rich motion control
Hybrid control architectures combining model-based and learning-based approaches
Comparative evaluations and benchmarking on real hardware
Practical engineering challenges: failure modes, interpretability, and maintainability
Future directions for learning-integrated locomotion control
For a more detailed description, please click the topics tab at the top of the page. If you believe you have late breaking results or would like to contribute to the discussion of these topics, please click the contribute button at the top of the page and submit your results by November 24th for potential inclusion in the poster session!