Model-based planning using GPU-accelerated Simulator as a World Model

Abstract

Manipulator robots are increasingly deployed in real-world tasks that require smooth, reactive motion and robust collision avoidance, particularly in dynamic and unstructured environments. This thesis presents a model-based, collision-free, online trajectory optimization framework tailored for such scenarios. The method involves sampling hundreds of trajectories from a multivariate normal distribution, shaping them with Bernstein polynomials, and evaluating them in parallel within a MuJoCo simulation. These trajectories are then optimized using the cross-entropy method. The system achieves real-time, in-the-loop planning by integrating a model predictive control strategy. The experiment, both in simulation and in real-world tests, demonstrated successful manipulation in an environment with multiple obstacles. In addition, the framework supports flexible task objectives by adjusting the cost function, enabling goal-driven behavior under varying conditions.