RACP: Risk-Aware Contingency Planning with Multi-Modal Predictions

TL;DR

RACP introduces a risk-aware contingency planning framework for autonomous driving that explicitly reasons about multi-modal, uncertain intentions of other road users. By embedding a Bayesian belief updater within a MoG-based prediction model, the approach generates a shared short-term plan up to a branching time $t_b$ and multiple contingent long-term plans conditioned on obstacle intents, all evaluated under a probabilistic risk metric that couples collision probability with severity. The method demonstrates improved efficiency and maintained safety across overtaking, urban T-junction, and intersection scenarios, outperforming robust, single-policy, and non-belief baselines, and scales to multi-vehicle scenes through permutation-based scene representations. The work provides a practical, closed-loop planning architecture with adjustable risk tolerance and online belief updates, indicating strong potential for real-world deployment and further research into risk metrics and prediction models.

Abstract

For an autonomous vehicle to operate reliably within real-world traffic scenarios, it is imperative to assess the repercussions of its prospective actions by anticipating the uncertain intentions exhibited by other participants in the traffic environment. Driven by the pronounced multi-modal nature of human driving behavior, this paper presents an approach that leverages Bayesian beliefs over the distribution of potential policies of other road users to construct a novel risk-aware probabilistic motion planning framework. In particular, we propose a novel contingency planner that outputs long-term contingent plans conditioned on multiple possible intents for other actors in the traffic scene. The Bayesian belief is incorporated into the optimization cost function to influence the behavior of the short-term plan based on the likelihood of other agents' policies. Furthermore, a probabilistic risk metric is employed to fine-tune the balance between efficiency and robustness. Through a series of closed-loop safety-critical simulated traffic scenarios shared with human-driven vehicles, we demonstrate the practical efficacy of our proposed approach that can handle multi-vehicle scenarios.

Figures (14)

• Figure 1: Two of the possible future intents of the human-driven vehicle are shown in red and blue. On the left, the ego-vehicle seeks a single plan which is safe with respect to both intents. On the right, a short-term trajectory is planned that branches into two contingent plans for each human's intents. The illustrative example is inspired by Urtasun.
• Figure 2: Illustration of the Frenet coordinate system.
• Figure 3: Snapshots from the CommonRoad simulation framework for an overtaking scenario where the ego-vehicle is depicted by the car icon and the obstacle-vehicle is illustrated as a blue rectangle. The obstacle-vehicle has three policies that it can execute, $\Lambda = \{\texttt{maintain speed}, \texttt{slow down}, \texttt{lane change}\}$, where the obstacle-vehicle's predicted trajectories are illustrated in distinct colors in \ref{['sub1']}. We first sample a set of short-term plans, $\mathcal{T}_{0:t_b}$ visualized in black, till the branching time $t_b$ shown in \ref{['sub1']}. We then sample a set of long-term plans, $\mathcal{T}_{t_b:T}$ depicted in purple, conditioned on the terminal states of the short-term plans, shown in \ref{['sub2']}. For the long-term plans, a cost is assigned to each plan per obstacle-vehicle policy $\lambda_i$. In Fig. \ref{['sub3']}, the total cost of the entire plan is computed by \ref{['10a']}, and the total optimal plan is the one given by $\tau^* = \tau^*_{0:t_b} \cup \{\tau^*_{t_b:T}(\lambda_1), \tau^*_{t_b:T}(\lambda_2), \tau^*_{t_b:T}(\lambda_3)\}$. In this example, the belief over the first policy is dominant, $b(\lambda_1) = 0.63$, compared to the rest causing the optimal short-term plan $\tau^*_{0:t_b}$ to be more biased towards $\tau^*_{t_{b:T}}(\lambda_1)$.
• Figure 4: Contingency planning paradigm. A large set of short-term plans are sampled together with multiple long-term plans. The total plan with the minimum cost for each possible future realization is selected (indicated in blue and red).
• Figure 5: The human-driven vehicle can have two different policies, lane-keeping (LK) or lane-change (LC). On the left, the ego-vehicle has a higher belief that the human-driven vehicle executes the LK policy. That's why the short-term trajectory tends to accelerate. On the right, the ego-vehicle has a higher belief that the human-driven vehicle executes the LC policy causing the short-term plan to decelerate and steer to the right.

Theorems & Definitions (7)

• Remark 1
• Definition 1: Risk Metric
• Remark 2
• Remark 3
• Remark 4
• Remark 5
• Remark 6