acceptodds
Withdrawn from ICLR 2027

STAR-Drive: Selective Trajectory Adoption and Replanning for Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision–language–action (VLA) models bring visual understanding and reasoning to autonomous driving, but their high inference cost makes frequent replanning expensive. In per-step replanning (Full Replan), each valid prediction immediately replaces the current driving plan, coupling the decision to request a new prediction with the decision to use it. However, the controller can continue following the current plan between model calls, and a newly generated plan need not be a better replacement. We introduce STAR-Drive, a training-free runtime that decides when to request a new plan and whether to adopt it, while leaving the VLA and feedback controller unchanged. Trajectory-guided scheduling uses the active trajectory’s age and upcoming motion to schedule model calls, limiting how long the plan can be reused and querying earlier before planned turns or speed changes. After each query, world-aware trajectory selection compares the candidate with the remaining active trajectory over the same future interval, checking for substantially reduced travel distance and vehicle footprints crossing mapped road boundaries. Rejecting a new plan does not extend the current plan’s reuse deadline; recovery rules govern execution when normal reuse is unavailable. Across 913 closed-loop AlpaSim scenes with Alpamayo 1.5, STAR-Drive increases the driving score, Score_AF, from 59.63 to 65.70 on the scale and reduces the off-road rate from 7.89% to 4.16% compared with Full Replan. It achieves these gains with 75.53% fewer VLA calls and a speedup in mean end-to-end latency, measured after initial rendering and excluding setup.

Primary area:
applications to robotics, autonomy, planning
Keywords:
autonomous driving, vision-language-action models, trajectory replanning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MNr2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STAR-Drive is a training-free runtime that decouples VLA querying from trajectory adoption for autonomous driving. It combines trajectory-guided scheduling, candidate selection based on path length and road geometry, and recovery rules, while leaving the underlying model and feedback controller unchanged.

Critical strengths
  1. The paper addresses a practically relevant problem in efficient VLA-based autonomous driving.
  2. The reported improvements in driving scores and inference efficiency are encouraging.
Critical weaknesses
  1. The proposed framework largely integrates existing ideas, including adaptive replanning, trajectory selection, and rule-based recovery. These components are primarily combined through heuristic mechanisms, without introducing sufficiently novel methodological insights. The contribution appears more like an engineering integration of existing techniques than a substantial research advancement.

  2. The evaluation is restricted to the AlpaSim simulator using a single VLA backbone. Experiments on more representative benchmarks, such as NAVSIM v2, which incorporates real-world driving data, are missing. This substantially limits the evidence for the method's generalizability and practical applicability.

  3. Although the paper reports reduced inference calls and wall-clock latency, the simulation does not account for actual inference delays during vehicle execution. More comprehensive real-time evaluations, including inference latency, control frequency, and deployment feasibility, are necessary to substantiate the claimed efficiency benefits in practical autonomous driving systems.

  4. The experimental comparisons are limited to a few adapted inference-time approaches, without sufficiently covering competitive SOTA methods for efficient VLA execution and autonomous driving planning. Stronger comparisons under matched computational budgets are needed to demonstrate the superiority of the proposed framework.

Questions for authors

The methodological novelty appears limited, as the framework primarily combines existing adaptive replanning, trajectory filtering, and rule-based recovery techniques. The authors should clarify the unique contribution and provide controlled ablations against simpler combinations of these components to demonstrate its added value. The experimental validation is insufficient, relying on a single VLA backbone and simulator. Stronger comparisons with relevant SOTA methods under matched computational budgets, along with evaluations on additional benchmarks such as NAVSIM v2, would better establish its generalizability.

Reviewer_jPxgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper points out that running a driving VLA in closed loop with Full Replan (querying the model at every control step of 0.1 s and switching immediately to the new prediction) couples two decisions: when to query and whether to adopt the returned prediction. STAR-Drive is a training-free execution rule that leaves the model and the controller unchanged. It queries the VLA only when the age of the current trajectory reaches 0.8 s or when large heading or acceleration changes lie ahead, and it compares each new candidate with the remaining active trajectory at matched future times, rejecting the candidate if its progress is less than half or if it crosses the HD-map road boundary. If rejection persists until the deadline, the system switches to a braking trajectory. On 913 AlpaSim scenes with Alpamayo 1.5, the paper reports ScoreAF 59.63 → 65.70, off-road rate 7.89% → 4.16%, 75.53% fewer VLA calls, and a 2.81× end-to-end speedup.

Critical strengths
  • This paper introduces a training-free runtime that coordinates inference scheduling ("when to query") and trajectory adoption ("whether to replace") through a single active trajectory. Leaving both the frozen VLA and the feedback controller untouched, it functions as an effective plug-and-play framework for trajectory-generating models.

  • It breaks the typical trade-off between efficiency and driving performance: reducing VLA calls by 75.53% (achieving a 2.81× end-to-end speedup) while simultaneously improving closed-loop driving quality Score_AF (59.63 -> 65.70); off-road rate (7.89% -> 4.16%).

Critical weaknesses
  • In Table 3 (100-scene development panel), enabling sparse querying alone without candidate checks raises Score_AF from 61.33 to 67.58, which accounts for over 58% of the total gain (61.33 → 72.06). A substantial portion of the performance improvement therefore appears to stem from open-loop receding-horizon chunk execution—a well-established property since ACT and Diffusion Policy—rather than the proposed trajectory-guided scheduling. While the authors compare against periodic execution on the 100-scene dev panel (72.06 vs. 66.15) and per-scene sweeps in Figure 4, the best fixed-interval configurations in Figure 4 actually outperform STAR-Drive in urban (65.55 vs. 58.26) and turning (91.70 vs. 86.24) scenarios. It remains unconvincing that the heuristic motion-demand triggers outperform a carefully tuned fixed execution interval. Comprehensive full-benchmark (913 scenes) comparisons against fixed-H chunk execution and temporal ensembling baselines are needed.

  • The central motivation is that driving VLAs suffer from high inference latency, yet the experimental setup explicitly ignores this latency by assuming queries complete synchronously and not imposing VLA execution time (1,591 ms per callback) as a delay in the simulator (p. 4, p. 7). In a physical deployment, with a 100 ms control step, an inference time of ~1.6 s spans roughly 16 control intervals, meaning a candidate cannot arrive within the nominal reuse deadline (H = 8, i.e., 0.8 s). Consequently, the reported 2.81× speedup reflects offline simulation throughput rather than closed-loop reaction responsiveness. Modern real-time frameworks address this latency through asynchronous planning while tracking the previous plan (e.g., Real-Time Chunking, NeurIPS 2025; AsyncDriver, ECCV 2024; DriveVLM-Dual, CoRL 2024). It is critical to demonstrate whether STAR-Drive maintains its advantages when realistic wall-clock execution delay is enforced.

  • The world-aware candidate check (Eq. 10) mirrors the evaluator's ground-truth OffRoad metric definition (Table 4) down to the 1 mm road-edge threshold, relying directly on the simulator's privileged HD map. None of the evaluated baselines (SP-VLA, SGAC, AAC) have access to this map. Crucially, Table 3 shows that adding the candidate check alone to per-callback replanning reduces off-road scenes from 6 to 2, fully matching STAR-Drive's final off-road performance. It is therefore difficult to isolate how much of the safety gain stems from principled trajectory management versus simply filtering candidates against the evaluator's exact metric definition using ground-truth map geometry. Comparisons against baselines equipped with the same geometric filter, as well as candidate filtering alternatives (e.g., Best-of-N selection or rule-based trajectory scoring in the spirit of PDM-Closed), are necessary. Furthermore, the Wrong Lane rate—which is not penalized by Score_AF—increases from 16.54% to 19.82%, raising concerns about unintended behavioral trade-offs.

Questions for authors

Please see the questions raised in the Weaknesses.

Reviewer_xwMzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Based on the fact that large VLA driving models' slow inference, the author proposed run-time training free decision mechanism to reduce the average latency while maintaining or slightly improving driving score on selected benchmarking dataset.

Critical strengths

Clear problem framing; training-free solution; informative ablations and diagnostics.

Critical weaknesses

Low practicality in actual product due to the worst-case latency will not improve; Driving score improvement may not reflect the real benefits as it is achieved by querying VLA less often.

Questions for authors

There seems to be many heavily hand-tuned thresholds. How do you come up with the current setting and how can we guarantee they are fit for general use cases?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.