Test-Time Trajectory Optimization for Autonomous Driving

summary

The gist

The gist: TOAD, a test-time procedure that turns the frozen scorer into the objective of a search over trajectories, at no training cost, improves results across six base planners on NAVSIM-v1 and

In short

TOAD is a test-time procedure that uses search to optimize trajectories based on a frozen scorer, without any training cost. It improves six base planners by searching for better paths that maximize the scorer's reward while maintaining driving comfort and safety. This method finds new, superior trajectories rather than just re-ranking existing ones.

Key concepts

TOAD: Test-Time Trajectory Optimization is a procedure that wraps an end-to-end planner in a search loop using the Cross-Entropy Method. It samples candidate trajectories, scores them with a fixed scorer, and refines the search around the best candidates. This allows planners to actively discover better paths during inference.
Frozen Scorer Generalization: A key finding is that not all trained scorers work as rewards; only those that remain accurate beyond their training proposals can be used effectively in TOAD. Scorers trained on a fixed vocabulary fail because they cannot explore new areas, whereas disentangled scorers that evaluate freely decoded trajectories perform better.
Optimization Objective J(u): The objective function for the search is defined as maximizing the learned reward S while penalizing deviations from an anchor trajectory and violating kinematic comfort limits. This ensures that the optimized trajectory is not only high-scoring but also dynamically feasible and comfortable, balancing performance with safety constraints.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Retrieval-Augmented Generation (RAG) — explained in 21 episodes
Sample Complexity — explained in 20 episodes
Flow Matching — explained in 19 episodes
Quantum Fisher Information (QFI) — explained in 19 episodes
Federated Learning — explained in 17 episodes

This episode discusses

The paper

Test-Time Trajectory Optimization for Autonomous Driving · Read on arXiv

Yihong Xu, Eloi Zablocki, Yuan Yin, Elias Ramzi, Ellington Kirby, Alexandre Boulch

valeo.ai · Sorbonne Universite

End-to-end planners for autonomous driving typically generate a set of candidate trajectories, score each one, and return the highest-scoring candidate. However, the scorer is applied only after the proposals are generated and cannot influence the set of trajectories: a weak set of candidates limits planning performance regardless of the scorer's quality. We instead treat the scorer as a learned trajectory-level reward function and search for trajectories that maximize it. Our method, TOAD, runs the Cross-Entropy Method at test time, warm-started from the planner's proposals. It requires no retraining and is plug-and-play for existing planners. Across seven base planners, TOAD improves results on NAVSIM-v1 (94.7 PDMS), NAVSIM-v2 (56.3 EPDMS), and the closed-loop HUGSIM benchmark. The code is made publicly available via the project page: https://valeoai.github.io/TOAD/.

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Hi, I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Test-Time Trajectory Optimization for Autonomous Driving".

Dev: Here's the gist of the paper. The gist: TOAD, a test-time procedure that turns the frozen scorer into the objective of a search over trajectories, at no training cost, improves results across six base planners on NAVSIM-v1 and NAVSIM-v2 while requiring only a frozen planner and a generalizing scorer.

Rosa: How it works TOAD wraps a frozen E2E planner in an inferencetime loop based on the Cross-Entropy Method (CEM)

1: .

Dev: At each iteration, the procedure samples trajectories from a Gaussian distribution, scores them with the frozen scorer, refits the Gaussian to the top-k candidates (‘elites’), and repeats

1: .

Rosa: The search is warm-started with the planner’s selected trajectory and confined to its neighborhood, which limits how far the search can drift from the scorer’s training distribution

3: .

Title and authors: Rosa: So we’re starting with the paper Test-Time Trajectory Optimization for Autonomous Driving. The title tells us immediately that this isn't about training a new model; it's about what happens during testing or when the system is already running.

Dev: Exactly. It focuses on optimizing the trajectory right when you are testing or running an autonomous system in real-time, instead of doing it all upfront during training <ref:2606.07170#pg1>.

Taro: The core idea is shifting the optimization from a pre-training step to an inference step where the system actively searches for better actions based on how well those actions score against a learned reward function <ref:2606.07185#pg3>.

Rosa: That means we’re taking the scorer, which is usually just a tool to rank options, and turning it into an active objective that the planner can manipulate during operation <ref:2606.07185#pg3>.

Dev: And the authors are bringing together a team with strong backgrounds in both planning and deep learning <ref:2606.07170#pg1>.

Taro: That mix is important because they're tackling the core challenge of end-to-end planning systems trying to generate safe and high-performing trajectories <ref:2606.07185#pg3>.

Rosa: The implication for us is that we can improve performance on existing, frozen planners right now without spending time retraining those large models <ref:2606.07170#pg1>.

Dev: They show how to attach a single frozen scorer to six different base planners like iPad, Hydra-MDP, GTRS, ZTRS, RAP, and DrivoR and immediately boost their performance on NAVSIM-v1 (ninety-four point seven PDMS) and NAVSIM-v2 (fifty-six point three EPDMS) <ref:2606.07170#pg2>.

Taro: That means researchers don't have to rebuild the entire planning pipeline just to see if a new search strategy helps them find a better path <ref:2606.07185#pg3>.

Rosa: It opens up a lot of possibilities for rapid prototyping and performance tuning in the autonomous driving research space because you can test new optimization methods on established architectures quickly <ref:2606.07185#pg3>.

Dev: The paper really is about exposing a hidden weakness in standard scorers that ranking evaluation cannot reveal because they are only checking a fixed set of candidates <ref:2606.07170#pg2>.

The paper's summary: Rosa: To summarize the core summary, Test-Time Trajectory Optimization for Autonomous Driving reframes the entire process by treating that scorer as a learned reward function and actively searching for trajectories that maximize it <ref:2606.07170#pg1>.

Dev: They detail how this search happens by running the Cross-Entropy Method at test time, where they sample controls from a Gaussian distribution, score them using the reward function along with comfort and anchor regularizers, and then update that Gaussian to focus on the best scores <ref:2606.07185#pg3>.

Taro: They also emphasize that this search is warm-started from the planner’s existing proposals, which keeps it focused near where the planner already thinks good paths exist <ref:2606.07185#pg3>.

Rosa: This approach is designed to be plug-and-play for existing planners because it only requires a frozen base planner and a generalizing scorer, making the optimization happen entirely during inference time <ref:2606.07170#pg1>.

Dev: The objective function they minimize includes not just the reward from the learned function S, but also penalties for deviating from an anchor trajectory using Canchor(u) and accumulating squared violations of standard kinematic comfort limits with Ccomf(u) <ref:2606.07185#pg3>.

Taro: So, they aren't just optimizing for a single metric; they are simultaneously ensuring the optimized path is also comfortable and stays close to the planner's original ideas <ref:2606.07185#pg3>.

Rosa: This dual optimization—maximizing the learned reward while respecting comfort constraints—is what allows them to achieve better results across six base planners on NAVSIM-v1, NAVSIM-v2, and HUGSIM <ref:2606.07170#pg1>.

Dev: The key finding they highlight is that this search itself is what drives the gains, not a stronger scorer; simply re-ranking the base planner’s proposals with the scorer can actually hurt performance <ref:2606.07185#pg2>.

The paper's improvements: Rosa: The authors suggest several key improvements to this methodology. First, they emphasize that the search is guided by the base planner’s uncertainty by setting the initial standard deviation of the Gaussian distribution from "the spread of the base planner’s own proposals" <ref:2606.07185#pg3>.

Dev: That means exploration naturally depends on where the base planner is already uncertain, making TOAD robust by guiding it toward areas that matter <ref:2606.07185#pg3>.

Taro: And they also show how to achieve state of the art performance on complex benchmarks like NAVSIM-v2 when using a DrivoR scorer because that scorer is disentangled from the trajectory generator, forcing it "to score freely decoded trajectories rather than a fixed grid" <ref:2606.07170#pg2>.

Rosa: That specific result showed gains as high as forty-three point six percent on the EPDMS metric for certain planners which is pretty significant when you look at the numbers <ref:2606.07185#pg2>.

Dev: They also show how this method improves DrivoR’s zero-shot HD-Score on the photorealistic closed-loop HUGSIM simulator <ref:2606.07185#pg3>.

Taro: The improvements also come from the fact that TOAD with the DrivoR scorer raises the safety sub-metrics jointly without collapsing progress which is a good thing for real driving systems <ref:2606.07185#pg3>.

Rosa: They also mention that their results show an increase in PDMS for all six base planners when focusing the search on boosting the ego progress sub-metric while maintaining on-par performance on safety metrics <ref:2606.07170#pg4>.

Dev: They also include two closed-form regularizers into the optimization objective: Canchor(u) equals u minus ubase squared, which ties the search to the trust region by penalizing deviation from the anchor, and Ccomf accumulates squared violations of standard kinematic comfort limits <ref:2606.07185#pg3>.

Conclusion: Rosa: So, to wrap up on this paper, Test-Time Trajectory Optimization for Autonomous Driving consistently outperforms the vanilla baseline across different hyperparameter settings and shows a low sensitivity to those choices <ref:2606.07170#pg4>.

Dev: The main implication is that the gain comes from the search discovering new trajectories rather than just re-ranking or smoothing existing ones, which is a really important distinction <ref:2606.07185#pg3>.

Taro: The paper's limitation is that warm-starting from a base planner’s proposals still exists, which makes researchers wonder if scaling the scorer with more data could eventually remove the need for a base planner entirely <ref:2606.07170#pg2>.

Rosa: And another limitation they flag is the cost associated with training that scorer, highlighting that future work should look at reducing reliance on costly trajectory-level annotations through self-supervised or world model-based rewards <ref:2606.07185#pg3>.

Dev: Overall, the paper on Test-Time Trajectory Optimization for Autonomous Driving shows a powerful way to improve planning performance by using test-time search guided by a learned reward function rather than just picking the best pre-generated candidate <ref:2606.07170#pg1>.

Taro: I just want to add that the method is plug-and-play, which is huge for real deployment because you don't have to rebuild your whole system when you tweak the optimization part <ref:2606.07170#pg1>.

Rosa: It’s definitely a solid piece of work showing how we can leverage test-time search to get tangible performance gains across six different driving environments <ref:2606.07170#pg4>.

More episodes

← Home