RGB: RL-guided MPPI gives humanoids precise whole-body control at 280 Hz

Researchers Yunsoo Seo et al. introduce RGB (RL Guided Whole-Body MPPI), a novel control framework that bridges the gap between the robustness of deep reinforcement learning and the flexibility of model predictive control. Instead of using a trained RL policy as the final controller, RGB uses it as a sampling prior that biases MPPI rollouts toward dynamically feasible behaviors. This allows the MPPI module to continuously correct the RL policy online to satisfy task objectives specified through modular cost terms—all without retraining the original policy. In simulations on a 29-degree-of-freedom Unitree G1 humanoid in MuJoCo, the system runs at an average of 280 Hz, demonstrating stable high-rate control. The method notably improves task-level precision over a pure RL baseline by correcting systematic drift during straight walking and tracking additional whole-body reference signals.

What makes RGB particularly compelling is its modularity and practical impact. Engineers can add new feedback objectives (e.g., changing foot placement or upper-body posture) simply by modifying the MPPI cost terms, rather than redesigning the RL reward function and retraining. The RL prior ensures sampling remains within dynamically feasible regions, preventing the MPPI from exploring unrealistic motions. This combination yields a controller that is both robust (from RL) and precise (from MPPI). The paper, submitted to arXiv, highlights a path toward more adaptable humanoid control that could accelerate real-world deployment in contact-rich environments like industrial assembly or disaster response. With humanoid robots entering the commercial spotlight, RGB offers a practical upgrade for existing control stacks.