New AI framework combines reinforcement learning with model predictive control to enable faster, more efficient robot decision-making at scale

Soft MPCritic merges RL and MPC approaches, learning value functions while using sample-based planning for both control and value target generation

The framework uses model predictive path integral control (MPPI) and trains a terminal Q-function with fitted value iteration to align learned values with the planner

An amortized warm-start strategy recycles previously planned action sequences to reduce computational costs while maintaining solution quality

The method uses an ensemble of dynamic models for scenario-based planning, making it practical for real-world applications requiring fast decision-making