PaperlayerView exact revision on arXiv

Source-linked AI summary

SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery Samples for Vision-Language-Action Model

Zewei Zhou, Ruining Yang, Xuewei, Qi, Yiluan Guo, Sherry X. Chen, Tao Feng, Kateryna Pistunova, Yishan Shen, Lili Su, Jiaqi Ma

arXiv:2604.19710v1cs.CV

TL;DR

Existing VLA driving models are limited by autoregressive action-generation latency and robustness gaps from predominantly positive-sample training. SpanVLA combines efficient flow-matching action bridging with GRPO-based negative-recovery training and a reasoning dataset, achieving state-of-the-art NAVSIM performance with reduced inference time.

  • Problem

    Existing VLA models face high action-generation latency from autoregressive decoding and limited robustness because they mainly learn from positive trajectories while overlooking negative and recovery behaviors.

  • Method

    SpanVLA bridges multi-granular VLM features to a flow-matching action expert conditioned on historical trajectories, then applies GRPO-based reinforcement fine-tuning with negative-recovery samples and introduces mReasoning.

  • Results

    SpanVLA achieves state-of-the-art performance across NAVSIM v1 and v2 while significantly reducing inference time.

  • Takeaways & Limitations

    Efficient action bridging and negative-recovery training improve the model’s planning performance and robustness within the evaluated NAVSIM benchmarks.

  • Takeaways & Limitations

    The current runtime is 1.5 Hz, which remains insufficient for direct real-world deployment without hardware-level or system-level acceleration.

Abstract

from arXiv · show

Vision-Language-Action (VLA) models offer a promising autonomous driving paradigm for leveraging world knowledge and reasoning capabilities, especially in long-tail scenarios. However, existing VLA models often struggle with the high latency in action generation using an autoregressive generation framework and exhibit limited robustness. In this paper, we propose SpanVLA, a novel end-to-end autonomous driving framework, integrating an autoregressive reasoning and a flow-matching action expert. First, SpanVLA introduces an efficient bridge to leverage the vision and reasoning guidance of VLM to efficiently plan future trajectories using a flow-matching policy conditioned on historical trajectory initialization, which significantly reduces inference time. Second, to further improve the performance and robustness of the SpanVLA model, we propose a GRPO-based post-training method to enable the VLA model not only to learn from positive driving samples but also to learn how to avoid the typical negative behaviors and learn recovery behaviors. We further introduce mReasoning, a new real-world driving reasoning dataset, focusing on complex, reasoning-demanding scenarios and negative-recovery samples. Extensive experiments on the NAVSIM (v1 and v2) demonstrate the competitive performance of the SpanVLA model. Additionally, the qualitative results across diverse scenarios highlight the planning performance and robustness of our model.

1 Introduction

Existing VLA driving systems face high autoregressive action-generation latency and limited robustness from learning mainly on positive trajectories. SpanVLA addresses these challenges with efficient action bridging, negative-recovery training, and evaluation showing strong NAVSIM performance.

  • End-to-end driving reduces modular error accumulation but conventional systems often lack environmental understanding and reasoning, especially in long-tail scenarios.
  • Existing VLA models face high autoregressive action-generation latency and limited robustness from relying primarily on positive or expert trajectories.
  • SpanVLA aggregates multi-granular sparse-layer VLM features and uses a flow-matching expert conditioned on historical trajectories to generate future actions efficiently.
  • GRPO-based reinforcement fine-tuning incorporates negative and recovery samples so the model learns to avoid undesirable behaviors and recover from challenging scenarios.
  • SpanVLA achieves state-of-the-art performance across NAVSIM v1 and v2 with significant inference-time reduction.

2 Related Work

Related work has explored VLM-guided driving, language-based action representations, and action experts to connect visual reasoning with continuous trajectories. SpanVLA builds on these directions with sparse multi-layer feature bridging, historical-trajectory initialization, and reinforcement fine-tuning using negative-recovery data.

  • VLMs are applied to autonomous driving because their reasoning capabilities and world knowledge can support continuous trajectory generation from visual observations and language instructions.
  • Some systems add VLMs for high-level guidance or supervision, but this can hinder full end-to-end optimization.
  • Action experts bridge VLM features to continuous trajectories, with prior work using final-layer features or dense-layer KV-caches for diffusion or flow-matching planners.
  • Existing reinforcement fine-tuning approaches generally rely on positive data, overlooking negative-recovery samples containing undesirable behaviors and recovery patterns.

3 SpanVLA

SpanVLA combines an autoregressive VLM backbone with an efficient flow-matching action bridge and trains it with reasoning, planning, and negative-recovery supervision. The bridge uses sparse VLM KV-cache features and historical trajectories to generate future actions, while reinforcement fine-tuning penalizes undesirable behavior and rewards expert recovery trajectories.

  • VLM Backbone: The VLM backbone jointly processes visual and language inputs to generate reasoning tokens and physical action tokens in an autoregressive sequence.During training, the model uses discrete action tokens to align reasoning with planning.
  • Efficient Action Bridging: The action bridge conditions on sparse VLM-layer KV caches and historical trajectory embeddings to generate continuous future trajectories through flow matching.A lightweight Transformer stack produces the conditioning, while the action expert predicts a vector field using historical trajectory and flow-time embeddings.
  • VLM Backbone: Adaptive reasoning switches between slow thinking with chain-of-thought and fast thinking with only action tokens.This mechanism avoids redundant reasoning in simple scenarios while retaining explicit reasoning for more complex cases.
  • Efficient Action Bridging: Historical trajectory initialization learns a transition from past actions to future actions instead of starting from pure Gaussian noise.Gaussian noise is added to historical embeddings during training to improve robustness, and optimal-transport interpolation defines the training path.
  • Reasoning and Negative-Recovery Data: mReasoning provides 30K expert-log samples from complex driving scenarios, while its negative-recovery subset contains 3K suboptimal trajectories and 3K corresponding expert corrections.The dataset covers interactions including lane changes, vulnerable road users, construction zones, and stop signs.
  • Reinforcement Fine-tuning: Reinforcement fine-tuning uses negative-behavior penalties for suboptimal trajectories and bounded recovery rewards for trajectories close to expert corrective actions.The recovery reward preserves flexibility by supporting multimodal recovery behaviors within a bounded L2 region.

4 Experiments

Experiments evaluate SpanVLA on NAVSIM benchmarks, action-generation efficiency, bridging choices, historical initialization, and RFT data and reward designs. Results report strong benchmark performance, reduced action-generation time, and improved planning from historical initialization and negative-recovery training.

  • NAVSIM Benchmark: SpanVLA outperforms state-of-the-art methods on NAVSIM v2 navtest and shows strong PDMS performance on NAVSIM v1 navtest.The evaluation covers NAVSIM v1 navtest, NAVSIM v2 navtest, and NAVSIM v2 navhard.
  • Efficient Action Bridging: The action bridge reduces overall action-generation time, with runtime growing only marginally as waypoint count increases.The efficiency comes from a lightweight action expert and parallel decoding, unlike autoregressive decoding whose time increases linearly with action length.
  • RFT and Negative-Recovery Samples: Mixing negative and recovery samples produces the best PDMS among the tested RFT settings on both benchmarks.The RFT schedule uses 2K positive warm-up steps followed by 4K mixed-training steps; moderate recovery inclusion performs best among tested mixtures.
  • Efficient Action Bridging: Flow matching achieves superior planning performance to the more efficient L1 action head.The comparison is made among action policies on NAVSIM v1 navtest.
  • Ablation Studies for Action Bridging: Historical initialization benefits driving performance, while sparser bridging-layer configurations improve efficiency but degrade driving performance.The full-caching baseline performs better than SpanVLA without historical initialization, whereas sparse configurations trade performance for efficiency.
  • Ablation Studies for RFT: Adding the L2 negative-behavior penalty improves performance with negative and recovery data, with best results at moderate weight and proximity threshold.The ablation evaluates the penalty term, its weight, and the threshold controlling when shaping activates.
  • Qualitative RFT Results: After RFT, negative-sample scenarios show more proactive merging and decisive lane changes, while recovery scenarios show successful constrained turns and returns to the correct lane.These qualitative comparisons are reported for mReasoning negative and recovery samples.

5 Conclusion

SpanVLA combines an efficient action bridge with flow-matching trajectory generation and GRPO-based learning from negative-recovery samples. Experiments on NAVSIM v1 and v2 demonstrate efficiency, robustness, and competitive performance.

  • Framework: SpanVLA integrates a VLM backbone, historical-trajectory-initialized flow matching, and an action bridge for future-trajectory generation.The framework uses VLM vision and reasoning guidance to generate trajectories efficiently.
  • Post-training: GRPO-based post-training uses negative-recovery samples to improve driving performance and robustness beyond positive driving demonstrations.The method targets both avoidance of typical negative behaviors and recovery behaviors.
  • Results: Experiments on NAVSIM v1 and v2 demonstrate SpanVLA's efficiency, robustness, and competitive performance.The conclusion reports these outcomes across the evaluated benchmarks.

A mReasoning Dataset

mReasoning addresses the lack of complex, long-tail driving scenarios and high-quality reasoning annotations by introducing a real-world dataset with automated CoT annotations and negative-recovery samples.

  • mReasoning targets complex, long-tail scenarios and negative-recovery samples that existing open-source datasets insufficiently cover.
  • The dataset is curated from in-house real-world driving logs collected by expert drivers.
  • The dataset includes 30K high-quality CoT annotations generated through an automated pipeline and checked by humans.
  • Its negative-recovery subset contains 3K suboptimal real-world ego trajectories and 3K expert corrections.

A.1 Reasoning Data

The reasoning-data pipeline combines diverse real-world driving scenarios, multimodal inputs, structured CoT prompting, automated generation, and human quality checks, while extending the dataset with negative-recovery samples.

  • Driving Scenarios: mReasoning collects 30K expert driving logs across complex scenarios, using a 1.5-second history horizon and a 5-second future horizon at 2 Hz.
  • Sensor Settings: Each sample includes eight surrounding camera streams, object annotations, and HD maps with lane, direction, and traffic-light information.
  • Reasoning Data: The annotation prompt defines a self-driving VLM role, CoT format, expert examples, and JSON outputs for critical components and driving decisions.
  • Reasoning Data: Inputs combine multiview cameras, ego states, high-level instructions, and ground-truth actions to support causal explanations linked to driving context.
  • Reasoning Data: Gemini-3-Pro generates reasoning data and checks format completeness and future-action alignment, while experts assessed 250 samples with 80.2% overall accuracy.
  • Negative-Recovery Data: The negative-recovery subset contains 3K suboptimal trajectories and 3K expert corrections, with labels distinguishing positive, negative, and recovery samples.
  • Evaluation Pipeline: The dataset extends NAVSIM with completed object annotations and HD maps to support PDMS-series evaluation.

B.1 Implementation Details

SpanVLA uses a two-stage training setup with a VLM backbone aligned to reasoning and planning, followed by action-expert fine-tuning and dual action representations.

  • Training: Stage one trains the VLM backbone on CoT reasoning data and ground-truth action tokens while freezing the Qwen2.5-VL-3B vision encoder.
  • Training: Stage two fine-tunes the action expert with the VLM backbone frozen and without action-token generation.
  • Action Representation: Future actions use continuous waypoints for flow matching and discrete action tokens for autoregressive decoding.
  • Action Codebook: The discrete action representation uses a codebook containing 2048 action tokens for 0.5-second vehicle motion.

C Details of RFT with Negative-Recovery Samples

The RFT pipeline combines driving-quality rewards with negative-behavior penalties, recovery rewards, reasoning regularization, action-reasoning alignment, and KL stabilization, while comparing action-bridging variants.

  • Driving Reward: PDMS and EPDMS provide the primary driving rewards for NAVSIM v1 and v2, respectively, measuring safety, compliance, comfort, and progress.
  • Driving Reward: EPDMS extends PDMS with driving-direction and traffic-light compliance plus lane-keeping, history-comfort, and extended-comfort metrics.
  • Negative-Recovery Reward: Negative samples subtract the reference-matching reward to penalize outputs near undesirable trajectories, whereas recovery samples add it to reward successful recovery behavior.
  • Negative-Recovery Reward: The reference-matching term equals 1 for an exact trajectory match, decreases linearly with deviation, and becomes 0 when deviation reaches δ.
  • Reasoning Regularization: The CoT penalty remains small below Ltol and increases smoothly as reasoning length exceeds that tolerance.
  • Reasoning Alignment: Action-reasoning alignment compares maneuver labels inferred from CoT keywords and trajectory geometry, penalizing inconsistent pairs.
  • Policy Regularization: KL regularization constrains the updated policy near the SFT reference policy, stabilizing updates while retaining SFT knowledge.
  • Action Bridging: Action bridging variants differ in whether the vector field is predicted directly in action space and in how VLM features are bridged.

D.2 Additional Ablation Studies for RFT

Additional ablations show that positive warm-up, balanced negative/recovery data, and calibrated reward shaping are important for effective RFT. Qualitative comparisons further illustrate improved handling of negative behaviors and recovery scenarios.

  • Data Recipe: The 2K warm-up + 3K positive + 0.5K negative + 0.5K recovery recipe achieves the best performance among tested data configurations.The total RFT budget is fixed at 6K samples, including 2K warm-up samples.
  • Warmup of RFT: Positive warm-up outperforms direct mixed training under both positive-plus-negative and positive-plus-recovery settings.The comparison keeps the total RFT budget fixed at 6K samples.
  • Negative Penalty and Recovery Reward: Removing the negative-behavior penalty and recovery-behavior reward lowers PDMS compared with enabling these terms.The ablation uses PDMS-only reward under positive, negative, and recovery data settings.
  • Negative Penalty and Recovery Reward: A moderate negative/recovery reward weight yields the best trade-off, while insufficient or excessive weighting degrades optimization or driving quality.Small weights provide insufficient shaping, whereas large weights can dominate the PDMS reward.
  • Negative Penalty and Recovery Reward: The proximity threshold controls when negative/recovery shaping activates, with smaller thresholds producing sparse supervision near reference trajectories.Activation depends on average L2 distance or ADE to the negative or recovery reference trajectory.
  • Qualitative Results: Negative samples reduce overly conservative stopping during turns or straight driving, while recovery samples support maneuvering when the intended lane is unavailable.The qualitative results compare planning behavior on negative and recovery samples before and after refinement.

E Limitation and Future Work

SpanVLA’s efficient action bridging accelerates inference, but the current implementation remains too slow for direct real-world deployment. Future work includes hardware and system-level acceleration and improved reward functions for negative-recovery data.

  • Runtime Limitation: The current runtime of 1.5 Hz remains insufficient for direct real-world deployment despite faster inference from efficient action bridging.The authors have not yet applied hardware-level optimization or deployment acceleration.
  • Future Deployment: The implementation requires 33 ms per token, while deployment-optimized systems can be substantially faster.The authors describe optimization strategies as orthogonal to SpanVLA and plan system-level acceleration.
  • Future Work: Designing more effective reward functions to fully exploit negative-recovery data remains an open problem.The paper identifies reward-function design as an important direction for future work.
Loading 2604.19710v1…