DBF-MA: A Differential Bayesian Filtering Planner for Multi-Agent Autonomous Racing Overtakes

TL;DR

The paper addresses overtaking in high-speed autonomous racing under opponent uncertainty. It introduces DBF-MA, a Bayesian trajectory inference framework that optimizes over a low-dimensional CBC parameter space to produce collision-free, dynamically feasible overtakes that start near the ORL and rejoin it after passing. The method combines a CBC-based trajectory parameterization with a three-component likelihood (collision avoidance, track-keeping, and dynamic feasibility) and performs Sequential Monte Carlo inference to construct a posterior p(\\theta | x) over overtaking maneuvers; finish-ahead and ORL rejoin are enforced within the sampling loop. Experimental results in Cavsim across three Formula 1 tracks show DBF-MA achieving 87% success, outperforming both a graph-based planner and Predictive Spliner, while maintaining lower DVS and CTE and offering favorable computation times, underscoring its practical potential for real-time, high-speed racing. Overall, the work presents a scalable, derivative-free, and risk-aware overtaking planner that integrates explicit track and tire dynamics constraints without restrictive footprint approximations, enabling safer and more aggressive autonomous racing strategies.

Abstract

A significant challenge in autonomous racing is to generate overtaking maneuvers. Racing agents must execute these maneuvers on complex racetracks with little room for error. Optimization techniques and graph-based methods have been proposed, but these methods often rely on oversimplified assumptions for collision-avoidance and dynamic constraints. In this work, we present an approach to trajectory synthesis based on an extension of the Differential Bayesian Filtering framework. Our approach for collision-free trajectory synthesis frames the problem as one of Bayesian Inference over the space of Composite Bezier Curves. Our method is derivative-free, does not require a spherical approximation of the vehicle footprint, linearization of constraints, or simplifying upper bounds on collision avoidance. We conduct a closed-loop analysis of DBF-MA and find it successfully overtakes an opponent in 87% of tested scenarios, outperforming existing methods in autonomous overtaking.

Figures (7)

• Figure 1: The Multi-Agent Differential Bayesian Filtering (DBF-MA) algorithm for overtaking trajectory synthesis. DBF-MA frames the problem as Bayesian inference over the parameters of a Composite Bézier Curve.
• Figure 2: The local planning problem on the Monza F1 Circuit. The goal of a local racing planner is to generate a trajectory that overtakes the target vehicle and returns to the Optimal Racing Line.
• Figure 3: Example ($N_S=3$) of a CBC trajectory $\mathcal{T}_\theta$. Black control points are fixed to satisfy initial and endpoint constraints; blue points are free parameters in $\theta$. Our approach guarantees that $\mathcal{T}_\theta$ will satisfy \ref{['eqn:initial_pos']} and \ref{['eqn:endpoint']}
• Figure 4: An illustration of $\Omega(\mathcal{T}_\theta(t))$
• Figure 5: A GG diagram showing $\mathbb{T}$ and $\Lambda$. The green acceleration vectors have a $\Lambda$ value of 0, while the red vectors have a strictly positive $\Lambda$.