Source-linked AI summary
DiffuSearch: How Hybrid Trajectory Planning Benefits from Aligned Objectives in Diffusion and Action Space
Steffen Hagedorn, Aron Distelzweig, Alexandru P. Condurache
TL;DR
Hybrid planners can become inconsistent when proposal generation and refinement optimize different objectives. DiffuSearch aligns these objectives across guided diffusion and MCTS, yielding strong closed-loop results with especially consistent gains from MCTS refinement and further benefits from objective sharing.
Problem
Hybrid planning modules may optimize different objectives, causing refinement to diverge from the initial proposal and produce suboptimal behavior.
Method
DiffuSearch generates a scene-consistent joint trajectory with guided diffusion, then locally refines it through MCTS in a discretized action space using shared driving objectives.
Results
DiffuSearch achieves competitive and state-of-the-art reactive closed-loop results across nuPlan and interPlan, with MCTS as the primary improvement source and further gains from shared objectives.
Takeaways & Limitations
Aligning implicit diffusion guidance with explicit MCTS rewards synchronizes trajectory generation and refinement, consistently reducing collisions and improving comfort in challenging interactive scenarios.
Takeaways & Limitations
The shared objective is used as a conservative coordination mechanism without claiming universal superiority over specialized modular designs.
Abstract
from arXiv · showhide
In trajectory planning for autonomous driving, hybrid planning architectures are often realized as a collection of disparate modules, each with its own objectives. This lack of a unifying principle can lead to inconsistencies between the initial and refined trajectory, resulting in suboptimal behavior. We address this by introducing DiffuSearch, a novel hybrid planner that uses a unified set of objectives across generation and refinement. Our model encourages all components to follow the same shared driving goals: collision avoidance, drivable area compliance, comfort, and progress. DiffuSearch employs a two-stage architecture. First, a guided diffusion model generates a scene-consistent, joint trajectory prediction, using our driving objectives as differentiable guidance functions to implicitly steer the denoising process. Second, a Monte Carlo Tree Search (MCTS) in a discretized action space performs an explicit, local refinement of this proposal, leveraging the same driving objectives as its reward function. This synergistic design leverages the diffusion model's strength in finding scene-consistent solutions combined with the explainable, constraint-aware refinement of MCTS. Experiments on nuPlan and interPlan reactive closed-loop benchmarks demonstrate that DiffuSearch achieves strong and often state-of-the-art performance, substantially reducing collisions and improving comfort, particularly in complex, interactive scenarios. Our ablation studies indicate that MCTS refinement is the main mechanism behind the gains, while sharing objectives between implicit guidance and explicit search provides further consistent improvements.
1 Introduction
Trajectory planning must balance flexible behavior with explainability and explicit safety checks. DiffuSearch addresses inconsistencies in hybrid planning by aligning objectives across proposal generation and refinement.
- Deep learning-based planners offer flexible behavior but limited explainability and no explicit safety checks by themselves.
- Hybrid planners add refinement stages, but disconnected module objectives can radically alter initial plans and produce suboptimal behavior.
- DiffuSearch combines a guided diffusion backbone with MCTS refinement in action space, using an initial joint prediction to constrain search.
- Using shared objectives for diffusion guidance and MCTS rewards provides additional gains beyond MCTS refinement alone.
2 Related Work
Prior work spans neural trajectory generation, diffusion-based planning, and MCTS refinement. DiffuSearch connects these paradigms by using shared objectives and a learned diffusion proposal for explicit local search.
- Monolithic neural planners flexibly adapt to traffic scenarios but lack explainability and safety guarantees.
- Diffusion models learn complex multimodal distributions through iterative denoising and can be controlled at inference through guidance.
- The guidance equation modifies the learned score using the gradient of an objective weighted by w.The displayed formulation is ŝθ(xt, t) = sθ(xt, t) − w ∇xt O(xt, t).
- Implicit diffusion guidance lacks formal guarantees of full constraint compliance, motivating explicit refinement with the same driving objectives.
- MCTS balances exploration and exploitation but benefits from a strong prior and discretized action space for focused, interpretable refinement.
3 Methodology
DiffuSearch combines guided diffusion with MCTS refinement under a unified set of driving objectives. The diffusion proposal supplies a focused prior, while action-space search explicitly refines it using interpretable rewards and reactive closed-loop evaluation.
- Overview: DiffuSearch combines a generative diffusion model with search-based refinement, using unified driving objectives across both stages.The architecture links scene-consistent joint prediction with explicit refinement in action space.
- Joint Prediction via Guided Diffusion: The diffusion model generates joint ego-and-agent trajectories conditioned on scene context and uses differentiable driving objectives to guide denoising.The generated ego trajectory becomes a strong prior that constrains subsequent search around the initial solution.
- Driving Objectives: The shared objective penalizes collision risk, drivable-area deviation, insufficient progress, and uncomfortable maneuvers.The objective terms are combined as collision, drivable, progress, and comfort components; safety receives the largest stated weight priority.
- MCTS in Action Space: MCTS explicitly refines the diffusion proposal in a discretized acceleration-and-steering action space using PUCT selection and a focused diffusion-derived prior.The search balances exploration and exploitation while concentrating actions near those prescribed by the diffusion trajectory.
- MCTS in Action Space: MCTS uses rewards with the same objective terms and weights as diffusion guidance, then retrieves a trajectory after a fixed computational budget.Simulation rewards are backpropagated to update visit counts and mean action values before selecting the highest-visit action sequence.
- Experimental Setup: Evaluation focuses on reactive closed-loop simulation in nuPlan, where scenarios run for 15 s at 10 Hz from real traffic situations.The ego vehicle is controlled by the planner and a low-level controller during simulation.
4 Experiments & Results
DiffuSearch performs strongly across reactive closed-loop benchmarks, with gains concentrated in difficult interactive scenarios. Ablations show that MCTS refinement drives most improvements, while aligned objectives and local search provide additional consistency and performance benefits.
- Benchmark performance: DiffuSearch achieves state-of-the-art results on all SMART-reactive benchmarks and two of three reactive nuPlan benchmarks.On Test14-hard-SR, it improves by up to 1.82% over DiffusionPlanner and 4.92% over PDM-Hybrid.
- Benchmark performance: In interactive SMART-reactive and harder Test14-hard scenarios, guidance and MCTS refinement help find safer, better local trajectory solutions.The refiner steers trajectories away from collision-prone interactions, smooths abrupt maneuvers, and can change lanes before critical pedestrian encounters.
- Ablations: Guidance alone decreases scores by up to −1.45% on Val14-R and −2.1% on Test14-SR, whereas MCTS alone improves scores by up to +1.11% on Test14-hard-SR.Combining guidance with MCTS achieves the best performance across all evaluated splits and benchmarks.
- Ablations: Aligned objectives synchronize trajectory generation and refinement, producing a more consistent planning process.The diffusion proposal places the plan in a promising solution region, while MCTS locally refines the same objectives rather than replacing it with a qualitatively different plan.
- Runtime trade-off: Planning performance improves until local refinement converges around 256 search steps, while MCTS runtime grows approximately linearly with search steps.Diffusion dominates runtime for small budgets, and 256 steps is used as the operating point in the experiments.
- Objective ablations: Each driving objective contributes to performance, but disabling MCTS reward components hurts more than disabling diffusion guidance objectives.Using the same objective set for diffusion guidance and MCTS reward yields the best observed planned trajectory.
- Driving metrics: DiffuSearch improves collision avoidance and time-to-collision while maintaining competitive drivable-area compliance and progress.Its objective weighting prioritizes collision avoidance, and critical scenarios may be handled by swerving into parking bays or onto the shoulder.
5 Conclusion & Future Work
DiffuSearch combines guided diffusion with MCTS refinement under a unified set of driving objectives. Across nuPlan and interPlan closed-loop benchmarks, this combination achieves strong performance, while shared objectives further improve collision reduction and ride comfort in challenging interactive scenarios.
- DiffuSearch combines a guided diffusion model with an MCTS refiner using a unified set of driving objectives.The objectives serve as implicit diffusion guidance and explicit MCTS rewards.
- The diffusion–MCTS combination achieves strong performance on nuPlan and interPlan closed-loop benchmarks.
- Sharing objectives across both planning stages consistently reduces collisions and improves ride comfort, particularly in challenging interactive scenarios.
- Future Work: Future work will add error feedback for temporal consistency and multimodal proposals for reasoning over distinct maneuvers in ambiguous scenarios.