RuleFuser: An Evidential Bayes Approach for Rule Injection in Imitation Learned Planners and Predictors for Robustness under Distribution Shifts

TL;DR

RuleFuser, an evidential framework, combines IL planners with classical rule-based planners to draw on the complementary benefits of both, thereby striking a balance between imitation and safety.

Abstract

Modern motion planners for autonomous driving frequently use imitation learning (IL) to draw from expert driving logs. Although IL benefits from its ability to glean nuanced and multi-modal human driving behaviors from large datasets, the resulting planners often struggle with out-of-distribution (OOD) scenarios and with traffic rule compliance. On the other hand, classical rule-based planners, by design, can generate safe traffic rule compliant behaviors while being robust to OOD scenarios, but these planners fail to capture nuances in agent-to-agent interactions and human drivers' intent. RuleFuser, an evidential framework, combines IL planners with classical rule-based planners to draw on the complementary benefits of both, thereby striking a balance between imitation and safety. Our approach, tested on the real-world nuPlan dataset, combines the IL planner's high performance in in-distribution (ID) scenarios with the rule-based planners' enhanced safety in out-of-distribution (OOD) scenarios, achieving a 38.43% average improvement on safety metrics over the IL planner without much detriment to imitation metrics in OOD scenarios.

Figures (5)

• Figure 1: RuleFuser adopts a parallel setup with two planners: a learned uncertainty-aware IL planner and a rule-based planner. (Top) For in-distribution driving data, in this illustration Boston, the IL planner learns to map the input to higher likelihood areas in latent space, indicating higher evidence. This gives the IL planner more pseudo-counts to contribute towards the Bayesian posterior contribution, surpressing the impact of the rule-based prior. (Bottom) For out-of-distribution scenarios, in this case Singapore, the input is mapped to lower evidence, so the posterior is largely controlled by the rule-based prior.
• Figure 2: Overview of RuleFuser. A spline generator produces a set of dynamically generated candidate future trajectories. Each one of these trajectories is augmented with a copy of scene context inputs (ego and agent histories, map and route information), and are processed individually by a transformer-based encoder. The regression head outputs an error trace for each candidate, while the classification head estimate the the likelihood of each candidate under the training data distribution. The results are fused with prior psuedo-counts from a rule-based planner to yield the final prediction.
• Figure 3: Figure shows the qualitative results for the three methods. While the predicted trajectories using Neural Predictor showcase good performance in Boston (ID), the performance deteriorates in the Singapore (OOD). Rule-aware Predictor has consistent performance in both ID and OOD but fails to capture nuances in speed and turning radius. RuleFuser shows consistent performance in both ID and OOD scenarios preferring higher performance in ID and falling back to safety in OOD.
• Figure 4: RuleFuser can generate an entire imitation-safety Pareto frontier, shown in this plot on combined ID and OOD test sets. The horizontal axis (safety score) decreases to the right, and the vertical axis (ADE) decreases upwards, so points further up and to the right are better. The bulge to the right indicates that mixing of the planners via RuleFuser results in better safety. RuleFuser also outperforms the nominal mixing model, indicating the importance of the evidential mixing strategy.
• Figure 5: Rear-end collision due to constant-velocity prediction. RH Planner plans the blue trajectories for the ego vehicle (pink box). Ground truth trajectory of non-ego vehicles (yellow box) is in green.