Better driving options do not always lead to better choices
ExecDiffuser: Closing the Sample-to-Execution Gap in Diffusion Planning · arXiv:2610.36837
In plain words
The paper introduces ExecDiffuser, a method for driving planners that generate possible paths and choose one for the vehicle. It trains path generation with evidence about how each path may perform, then learns to choose a path that passes an online validity check.
That chosen trajectory becomes context for the next planning round. If none pass, it invokes a deterministic minimal-risk maneuver.
In nuPlan, a simulated-driving benchmark that scores planners in closed-loop runs, ExecDiffuser scored higher on average than two reproduced planners across six settings. A separate candidate test showed that improving the available paths alone did not ensure the chosen path scored better.
The research
If a driving planner can generate a strong path, will it actually choose it? In closed-loop driving, the vehicle follows the published trajectory and moves into a changed scene before the next planning cycle. Candidate quality alone may therefore miss whether the planner’s committed decision is good.
ExecDiffuser trains trajectory proposals with execution evidence and learns a recurrent selector to publish a candidate that passes its validity check, returning it as context for the next cycle; if none pass, it invokes a deterministic minimal-risk maneuver.
nuPlan tests planners in closed-loop driving simulations and scores their performance. Across its six settings, ExecDiffuser averaged 85.62 points over five seeds, above Diffusion Planner (81.60) and FlowDrive (84.31), both reproduced under the same protocol.
On matched Val14 states, proposal-only training improved the best candidate’s score but slightly lowered the score chosen by the same imitation rule. This shows why evaluating candidates alone can miss the decision that controls the next driving state.
The evaluations use simulators and a separate driving test; the paper reports no real-world interactive validation. A diagnostic also found false-valid decisions at occluded intersections, so these results do not establish reliable validity checks in every scene.
What the evidence shows
- The candidate and the choice
ExecDiffuser is a recurrent driving-planning method: it generates candidate trajectories and publishes one admitted by its online validity check to control. If none pass, it invokes a deterministic minimal-risk maneuver.
When it publishes a candidate, that trajectory feeds into the next planning cycle, linking generation, commitment and replanning. The paper asks whether evaluating candidates alone captures the decision the vehicle will follow.
It uses nuPlan, a simulated-driving test that scores planners in closed-loop runs, alongside matched diagnostics to compare available options with the published choice.
- When options outpace choice
To separate available options from the choice, the authors cloned eligible states in Val14, one nuPlan evaluation setting, with at least one valid candidate. A simulator evaluator scored each trajectory in an independent eight-second run without replanning; the authors calculated the best score in the complete set and the published candidate’s score, then averaged across five seeds.
The best included candidates rejected by the online validity check. With the same imitation-based choice rule, proposal-only training raised the best score from 90.33 to 91.36 points but lowered the chosen score from 89.82 to 89.64 versus base proposals, widening the score gap.
- Improving what can be chosen
Candidate generation itself has a separate training change. Intent-Guided Diffusion Proposal (IGD) assigns training responsibility using closeness to demonstrated paths and execution evidence, then a separate shaping step adjusts trajectory geometry.
With the full Execution-Grounded Selection (EGS) component held fixed, adding execution-conditioned responsibility and shaping raised the complete-set best score from 91.36 to 92.63 nuPlan points, averaged across five seeds. That higher ceiling does not say what the planner actually publishes.
- Turning options into decisions
Execution-Grounded Selection (EGS) addresses the choice stage. It filters invalid proposals, learns a context-dependent utility for ranking those that pass, and uses scene and execution evidence plus the previous publication.
The chosen path returns as context for the next cycle; if none passes, the planner uses a deterministic minimal-risk maneuver. With the proposal generator fixed, a matched Val14 comparison found EGS selected the best-scoring candidate in a larger share of eligible states and reduced the average score gap to the retrospective best versus imitation selection.
- Testing the full driving loop
Across six nuPlan settings, ExecDiffuser averaged 85.62 points over five seeds, above reproduced Diffusion Planner (81.60) and FlowDrive (84.31) under a shared version, scenario, simulator and scoring protocol. In a separate matched intervention, the same frozen planner scored 1.90 points higher over an eight-second run after an EGS-selected four-second opening than after an imitation-selected one.
These results support the authors’ view that selection can matter after the vehicle executes a trajectory.
Deep
- Original paper authors
- Shuo Pei, Bosun Liang, Saihui Zha, Chen Sun, Huachun Tan
- Submitted to arXiv
- License
- arXiv non-exclusive