From Prediction to Planning With Goal Conditioned Lane Graph Traversals

TL;DR

This work tackles turning trajectory prediction advances into goal-directed planning by introducing route-conditioned lane-graph traversals. By conditioning prediction on a navigation goal at the behaviour level, the authors convert a multimodal predictor (PGP) into a goal-conditioned planner (GC-PGP) that constrains traversals to route-compliant paths while preserving multi-horizon reasoning. They implement soft and hard route masking, evaluate on the nuPlan dataset with open- and closed-loop simulations, and show that GC-PGP substantially narrows the gap to state-of-the-art planning baselines and improves route adherence and interpretability. The results indicate that goal-conditioned prediction can serve as a strong planning baseline and a practical starting point for developing more robust, interpretable planning systems for autonomous driving.

Abstract

The field of motion prediction for automated driving has seen tremendous progress recently, bearing ever-more mighty neural network architectures. Leveraging these powerful models bears great potential for the closely related planning task. In this letter we propose a novel goal-conditioning method and show its potential to transform a state-of-the-art prediction model into a goal-directed planner. Our key insight is that conditioning prediction on a navigation goal at the behaviour level outperforms other widely adopted methods, with the additional benefit of increased model interpretability. We train our model on a large open-source dataset and show promising performance in a comprehensive benchmark.

Figures (5)

• Figure 1: Left: Candidate plans on an unconditioned graph cover all feasible behaviours, including those that do not follow the route. Right: Goal-conditioning is achieved by excluding trajectories that traverse off-route graph nodes (grey) from being considered. Hence, the model can use its entire multimodality budget to find trajectories that comply with a navigation goal, potentially yielding better plans.
• Figure 2: Method overview: We manipulate the edge probabilities of a graph-based scene representation. Route-conditioned trajectories are then decoded from the manipulated graph and the most likely one selected as a plan for the SDV.
• Figure 3: Qualitative results: Left to right depicts the first 4 seconds of the SDV making a left turn. Top: GC-PGP, Bottom: node features ablation. The expert trajectory and model prediction are depicted in orange and blue respectively. The orange surface indicates parking areas. Stop lines are shown in red. In the bottom row toggling between turning and going straight manoeuvres can be observed, whereas our method (top) results in plans that are compliant with the route at all times. (Best viewed in colour.)
• Figure 4: Making a turn in closed-loop simulation. The SDV enters the intersection after 3 seconds having travelled $7$% of the expert's total travelled distance and leaves the intersection after 6.7 seconds with $26$% progress. After 15 seconds of simulation, it covers $64$% of the expert's path. (Best viewed in colour.)
• Figure 5: Overview of the PGP Model. Adapted from deo2022multimodal. The model consists of three modules. First, the encoder takes the history of surrounding agents as well as the SDV motion and map information as input and encodes with Gated Recurrent Units (GRU). Agent-Node Attention and GNN layers generate a road graph containing real-time traffic information. Subsequently, the policy header predicts probabilities for the outgoing edges of each graph node. Sampling these probabilities yields traversals, that describe a likely future behaviour. Finally, the latent-variable decoder regresses the exact motion trajectory based on the traversals.