Transformer-Based Model Predictive Path Integral Control

TL;DR

This paper introduces TransformerMPPI, a transformer-based initialization method for Model Predictive Path Integral control, significantly enhancing sample efficiency and computational speed in complex control tasks by leveraging long-horizon pattern recognition.

Contribution

The paper proposes a novel transformer-based approach to initialize MPPI control sequences, improving efficiency and performance over traditional methods.

Findings

TransformerMPPI outperforms traditional MPPI in average cost reduction.

It achieves higher sample efficiency in control tasks.

It demonstrates faster computation in obstacle avoidance scenarios.

Abstract

This paper presents a novel approach to improve the Model Predictive Path Integral (MPPI) control by using a transformer to initialize the mean control sequence. Traditional MPPI methods often struggle with sample efficiency and computational costs due to suboptimal initial rollouts. We propose TransformerMPPI, which uses a transformer trained on historical control data to generate informed initial mean control sequences. TransformerMPPI combines the strengths of the attention mechanism in transformers and sampling-based control, leading to improved computational performance and sample efficiency. The ability of the transformer to capture long-horizon patterns in optimal control sequences allows TransformerMPPI to start from a more informed control sequence, reducing the number of samples required, and accelerating convergence to optimal control sequence. We evaluate our method on various control tasks, including avoidance of collisions in a 2D environment and autonomous racing in the presence of static and dynamic obstacles. Numerical simulations demonstrate that TransformerMPPI consistently outperforms traditional MPPI algorithms in terms of overall average cost, sample efficiency, and computational speed in the presence of static and dynamic obstacles.