EventDrive: Event Cameras for Vision-Language Driving Intelligence

A new benchmark and model suite called EventDrive aims to integrate event cameras into vision-language systems for autonomous driving, unifying asynchronous brightness-change data with RGB frames and language supervision across four core dimensions [1]. Event cameras capture brightness changes asynchronously with microsecond latency and high dynamic range, providing motion fidelity that conventional frame-based sensors often miss [1]. These properties make them a candidate complement to RGB in autonomous driving, particularly under blur, glare, and rapid motion where frame-based perception can become unreliable [1]. However, existing event-aware vision-language models have been limited to generic perception tasks and have not addressed how event sensing contributes to reasoning and decision-making across the full driving loop [1]. EventDrive addresses this gap by unifying event streams, RGB frames, and language supervision across four dimensions: Perception, Understanding, Prediction, and Planning [1]. The benchmark covers captions, structured question-answering, grounding, motion-state recognition, trajectory forecasting, and planning tasks [1]. The accompanying model, EventDrive-VLM, introduces a multi-horizon event pyramid and a temporal-horizon mixture-of-experts module to adaptively encode and fuse asynchronous and frame-based information for downstream reasoning [1]. Comprehensive evaluation across diverse tasks shows that event streams provide substantial gains in temporal precision, motion awareness, and robustness [1]. The work positions event sensing as a central component of driving intelligence rather than a peripheral sensor modality [1]. The approach aligns with broader efforts to build internal representations of environments, known as world models, which predict how environments change over time in response to actions [3]. World models help agents plan, reason, and act without constant real-world trial and error, simulating dynamics such as physics, object interactions, and causality [3]. Modern versions power robots, autonomous driving, and interactive video generation [3]. Autonomous driving systems remain constrained by software and mapping requirements needed to operate safely across the wide variety of conditions drivers experience [5]. As of late 2025, no system has achieved full autonomy in all domains, sometimes referred to as Level 5 on the SAE International scale of 0 to 5 levels of automation [5]. The primary technologies used include LiDAR and visual sensors such as cameras, combined with GPS, neural networks, and artificial intelligence [5]. Event cameras offer a sensing modality that captures temporal structure conventional exposures often miss, potentially addressing some of these software challenges [1]. Large language models, which underpin modern vision-language systems, are typically based on transformer architectures and are pre-trained to predict the next word before being fine-tuned for specific tasks [4]. Benchmark evaluations for such models attempt to measure reasoning, factual accuracy, alignment, and safety [4]. EventDrive extends this paradigm by incorporating event-camera data into the evaluation framework for driving-specific reasoning [1].