Sampling-based motion planning has gained traction in robotics, and the Model Predictive Path Integral (MPPI) controller is a notable example. Its appeal lies in handling nonlinear dynamics and non-smooth objectives without requiring gradient information, making it flexible for complex tasks. However, execution delay—the lag between when a plan is computed and when it is actually applied—can degrade performance in real-time systems.
The new method, CLIS-MPPI, targets this problem by adding initial-state sampling within a closed-loop framework. Rather than assuming the planned trajectory starts immediately, it samples possible initial states that account for the delay, allowing the controller to anticipate and compensate for the lag. This extends the standard MPPI approach to more realistic operating conditions.
Because the method is closed-loop, it can incorporate feedback as the robot moves, refining its plan based on the latest state. The paper's abstract highlights this as a way to maintain the benefits of MPPI while addressing a practical challenge in deployment. The details of the algorithm and its evaluation are not fully described in the available abstract, but the contribution is clearly positioned as an improvement for delay-prone robotic systems.