Jul 2026· Annual Conference on Genetic and Evolutionary Computation· pp. 1038-1046· 0 citations· 34 references
Computer Science
TL;DR
This work creates a novel open-source symbolic traffic model EvoDrive designed specifically for EC research, which outputs LLM-readable snapshots and shows that LLMs + EvoDrive with CLEAR can reduce error by more than 20% compared to without CLEAR, with statistically significant results.
Abstract
The overtake car maneuver involves high risk and complex judgement. For autonomous vehicles this is challenging, especially for human-initiated overtake requests. If a user requests the maneuver there must be a rapid safety assessment. Language models have great potential to classify safety with explanations, but they struggle to disentangle critical information from complex vehicular environments. We apply CLEAR (Cue Learning using Evolution for Accurate Recognition) to evolve prompt cues that optimize the ability of language models to correctly predict safety scores for overtaking maneuvers. To achieve this, we create a novel open-source symbolic traffic model EvoDrive designed specifically for EC research, which outputs LLM-readable snapshots. We show that LLMs + EvoDrive with CLEAR can reduce error by more than 20% compared to without CLEAR, with statistically significant results. Analysis shows evolved cues are coherent and have reduced variability in LLM output.
Autonomous robot navigation failures differ not only in categorical severity but also in the physical context in which they occur. A near-miss at low speed under reliable sensing is not equivalent to the same event during rapid motion, close obstacle approach or degraded perception. This paper reframes navigation failure prediction as consequence-sensitive forecasting. We first establish a fixed baseline in which training weights are modulated by categorical severity, then introduce an adaptive extension defining a state-dependent consequence function combining severity with normalised velocity, obstacle proximity and sensing uncertainty, together with a risk-sensitivity term that rises as conditions deteriorate. We evaluate on 2,000 simulated differential-drive episodes (~1,000,000 timesteps) using episode-level GroupKFold, with external validation on the UCI SCITOS G5 dataset. Fixed weighting raises Logistic Regression high-severity recall from 0.851 to 0.985 and reduces missed consequence cost from 1,940 to 313; the adaptive extension reaches 0.998 and 82. Under matched false-positive conditions, however, the discriminative advantage is modest (0.986 versus 0.984), so most of the gain reflects a more conservative operating point rather than better ranking. The effect is consistent across all five folds and stable across a threefold span of context coefficients. Because the primary simulation produced no collisions, we add a controlled extension in which 108 of 600 episodes terminate in contact: collision recall rises from 0.850 to 0.966 (fixed) and 0.984 (adaptive), with missed collision cost falling from 1,000 to 105, at false-positive rates of 0.413 and 0.799, respectively. Context-dependent consequence modelling thus provides a principled mechanism for allocating conservatism by physical risk.
Vision-language models (VLMs) have recently emerged as a promising paradigm for end-to-end autonomous driving, enabling agents to map multimodal inputs and high-level navigation instructions directly to executable trajectories. However, in practice, these models exhibit a persistent command-following gap: predicted trajectories often show weak sensitivity to navigation commands, resulting in incorrect behavior at critical decision points. We identify this issue as a form of conditional policy collapse, where regression-based training under multimodal trajectory distributions encourages the model to rely on dominant visual priors while marginalizing the language-conditioned signal. To address this issue, we introduce a principled formulation of classifier-free guidance (CFG) for regression-based vision-language driving. We show that CFG can be interpreted as isolating the instruction-induced residual in the action space by contrasting conditional and unconditional predictions, thereby explicitly amplifying the effect of the navigation command at inference time. However, a standard two-pass CFG introduces prohibitive latency for real-time control and produces noisy instance-level guidance directions. Building on a mean-shift interpretation of CFG, we propose Latent-Centroid Steering (LCS), a single-pass guidance mechanism that replaces instance-level residuals with class-level latent shifts. By projecting conditional representations toward precomputed command-specific centroids, LCS performs class-level latent steering based on cluster geometry that is both more stable and computationally efficient. We demonstrate that LCS reduces inference latency by approximately 50% while achieving stronger command adherence and improved driving performance on both closed-loop (Bench2Drive) and open-loop (nuScenes) benchmarks. Code will be released.
Meibo Hu, Jiamian Wang, Pichao Wang et al.· 0 citations
Findings characterize model-level behavioral priors relevant to LLM choice and prompt design in AV applications, and show that LLM decisions reflect a mix of model characteristics, linguistic framing, and scenario context.
Zhipeng Bao, Wen-Jie Zhao, Qianwen Li· Journal of Intelligent and C...· 0 citations
Safe quadrotor navigation in cluttered and dynamic environments depends not only on instantaneous geometric perception, but more critically on anticipating collision risks induced by relative motion. Conventional modular pipelines frequently suffer from perception latency, while end-to-end learning methods relying on implicit scalar rewards often struggle to extract reliable spatio-temporal features without physics-grounded supervision. To address this, we propose an anticipatory risk-guided reinforcement learning framework. Leveraging privileged simulator states, we construct a directionally aligned future collision risk map based on the Closest Point of Approach (CPA). Through an asymmetric actor-critic architecture, the network is trained to self-predict this structured risk, which explicitly guides the visual policy during deployment. A lightweight spatio-temporal encoder extracts motion cues directly from onboard depth sequences, bypassing explicit object tracking or optical flow estimation. Extensive simulated and real-world experiments demonstrate that our method effectively improves safety margins and flight efficiency in dense dynamic clutters compared to existing baselines. Furthermore, the learned policy achieves robust zero-shot Sim-to-Real transfer on a physical quadrotor, relying purely on abstracted spatio-temporal depth sequences and its self-predicted risk priors, validating the effectiveness of our approach and its robust generalization from simulation to reality.
Yuchao Mei, Guohao Zhang, Luxia Ai et al.· arXiv.org· 0 citations
Evaluating large language models (LLMs) in safety-critical, physics-governed environments requires more than accuracy-based metrics, because predictions that are numerically close to the ground truth can still violate operational constraints, combine fields in physically inconsistent ways, or fail to produce usable structured outputs. Existing evaluation protocols do not measure these failure modes reliably. We propose FLY-EVAL++, an evidence-driven evaluation protocol that combines deterministic verification of protocol compliance, physical feasibility, and safety constraints with fixed rubric-guided aggregation into interpretable multi-dimensional scores. We instantiate FLY-EVAL++ for Flight Trajectory and Attitude Prediction (FTAP) by extending the PilotBench setting with history-conditioned and multi-step prediction tasks. Across 66 LLMs, safety compliance is the most discriminative dimension of model behavior: models with comparable predictive performance differ by more than 28 points in safety score, and we observe recurrent failures including safety violations under physically plausible predictions and instability in multi-step rollouts. These results show that evaluation in safety-critical domains should measure constraint satisfaction and structured validity explicitly rather than rely on accuracy-centric reporting alone.
Ya-Lun Wu, Jun-Feng Fang, Jia-Wei Wang et al.· 0 citations
Accurately predicting pedestrian crossing intent is paramount for the safety and reliability of autonomous driving systems in complex urban environments. Despite tremendous progress in the past decade, existing methods often fail to make reliable predictions from instantaneous observations. To address this limitation, we propose a novel privileged knowledge distillation (PKD) framework that can be seamlessly integrated into existing prediction models to improve their performance under instantaneous observation scenarios. Specifically, we propose a unique training paradigm where a privileged branch (teacher) learns temporal trend features from bidirectional privileged sequences, encompassing both extended historical observation sequences and future behavioral sequences in the time-to-event (TTE) period. This bidirectional trend knowledge is then distilled into the instantaneous branch (student) via a meticulously designed composite loss, enabling the student to effectively simulate and infer complex pedestrian intent from only a few frames of input at test time. The framework exploits specialized components, including a Gated Multi-scale Trend Module (GMTM) and a bidirectional trend simulation module (BTSM), to capture and mimic privileged knowledge explicitly. Extensive experiments on the public PIE and JAAD datasets demonstrate that our PKD framework significantly enhances the performance of multiple state-of-the-art baseline models, confirming its effectiveness and broad applicability. Note to Practitioners—In real-world autonomous driving, vehicles often need to predict a pedestrian’s intent to cross the road using only a few visual frames, especially in sudden or complex traffic situations. However, short and incomplete observations frequently lead to unreliable intent predictions. This study aims to improve the reliability of such instantaneous pedestrian intent prediction when limited sensory data are available. We propose a Privileged Knowledge Distillation (PKD) framework that is flexible and can be plugged into existing prediction systems without additional sensing hardware. During training, the privileged branch learns from both extended past and future sequences to form a deeper understanding of how pedestrians’ intent evolves over time. The knowledge of the privileged branch is distilled into an instantaneous branch that only relies on short observations. At runtime, the instantaneous branch can reason about pedestrians’ intent using only brief observations. In practice, this approach can enhance the safety and responsiveness of autonomous vehicles by allowing them to make early and reliable decisions even with minimal visual input. Therefore, our approach has broad application prospects in autonomous driving and automatic robotics.
