HYPE: Hybrid Planning with Ego Proposal-Conditioned Predictions

TL;DR

HYPE addresses safe, interpretable motion planning in urban multi-agent environments by explicitly modeling bidirectional ego–other interactions. It introduces an architecture that fuses multimodal ego trajectory proposals with Monte Carlo Tree Search refinement and an ego-conditioned occupancy predictor, enabling interaction-aware planning with minimal handcrafted costs. The approach uses a planning horizon of $3$ seconds and a $400×400$ occupancy grid, and demonstrates state-of-the-art safety and adaptability on nuPlan and DeepUrban benchmarks, supported by extensive ablations. Overall, HYPE offers a robust, interpretable framework that leverages learned proposals to constrain search while maintaining safety guarantees in complex urban driving.

Abstract

Safe and interpretable motion planning in complex urban environments needs to reason about bidirectional multi-agent interactions. This reasoning requires to estimate the costs of potential ego driving maneuvers. Many existing planners generate initial trajectories with sampling-based methods and refine them by optimizing on learned predictions of future environment states, which requires a cost function that encodes the desired vehicle behavior. Designing such a cost function can be very challenging, especially if a wide range of complex urban scenarios has to be considered. We propose HYPE: HYbrid Planning with Ego proposal-conditioned predictions, a planner that integrates multimodal trajectory proposals from a learned proposal model as heuristic priors into a Monte Carlo Tree Search (MCTS) refinement. To model bidirectional interactions, we introduce an ego-conditioned occupancy prediction model, enabling consistent, scene-aware reasoning. Our design significantly simplifies cost function design in refinement by considering proposal-driven guidance, requiring only minimalistic grid-based cost terms. Evaluations on large-scale real-world benchmarks nuPlan and DeepUrban show that HYPE effectively achieves state-of-the-art performance, especially in safety and adaptability.

Figures (7)

• Figure 1: Different proposal strategies in the hybrid planners. (a) Existing sampling proposals are subject to a given optimization horizon. (b) Our proposed approach employs a learnable multimodal ego proposal network to generate flexible trajectories that capture long-term scene interactions.
• Figure 2: Model architecture of the hybrid heuristic planner HYPE. The model integrates an ego proposal network that explicitly encodes route embeddings to generate multimodal ego future proposals, which condition the occupancy prediction and guide the subsequent MCTS refinement.
• Figure 3: Proposal-guided MCTS planning. During expansion, nodes are chosen from the set of perturbed proposal candidates. In simulation, each rollout is evaluated using a lightweight cost function combining collision risk and proposal deviation.
• Figure 4: Qualitative results for ego-conditioned occupancy prediction. Our model predicts occupancy grids for surrounding agents conditioned on different ego trajectory proposals. In the visualization, the ego vehicle is shown in yellow, and its trajectory proposals are in red. Occupancy predictions of other agents at future timesteps are color-coded: red for $t=1s$, green for $t=2s$, and blue for $t=3s$. White regions indicate similar occupancy probabilities across all timesteps, typically reflecting static agents.
• Figure 5: Qualitative results of 8-second trajectory planning in a crowded urban scenario from DeepUrban MunichTal. Each subfigure visualizes the planned ego trajectory from different planners: HYPE (ours), DTPP, and DiffStack. The red line represents the planned future trajectory of the ego vehicle, with the yellow box indicating its current position. Magenta boxes denote surrounding agents along with their ground-truth future trajectories.