自动驾驶车辆在道路上面对多种动态风险，而一项本周发表在《Nature Communications》的研究提出受人眼启发的人工视觉系统，速度达到当前最先进水平的4倍。该进展旨在把对复杂环境的理解从“慢于人类”推向“接近或超过人类反应”的时间尺度。

主流的光流方法需要处理每一帧的每个像素，计算代价高；即使在最先进技术下，在单帧中区分不同物体也可能耗时超过0.6秒。对以高速公路速度行驶的自动驾驶汽车而言，每0.5秒的延迟会带来约12米的“过时信息”行驶距离，放大了安全与控制风险。

研究通过在神经形态硬件上加入类似外侧膝状体核（LGN）的注意力过滤层来优先处理时空变化区域，在多种任务中实现约4倍速度提升且保持或提高精度。对自动驾驶场景提升尤为显著，准确率翻倍，并且在多数情况下速度超过人类水平，但在复杂、密集运动场景中精度会下降，且仍需回馈到传统算法因而受其短板限制。

Autonomous vehicles face many dynamic road hazards, and a study published this week in Nature Communications reports a human-eye-inspired artificial-vision system that runs 4x faster than the current state of the art. The goal is to shift understanding of complex scenes from “slower than humans” toward time scales that match or exceed human responses.

Optical-flow methods typically process every pixel in every frame, making them computationally intensive; even with state-of-the-art technology, distinguishing different objects in a single frame can take over 0.6 seconds. For an autonomous vehicle at motorway speeds, each 0.5-second delay corresponds to about 12 metres traveled using outdated information, compounding safety and control risk.

The work adds an LGN-like attention-filter layer on neuromorphic hardware to prioritize regions with spatiotemporal change, delivering roughly 4x speed while maintaining or improving accuracy across tasks. Gains are strongest in autonomous driving, where accuracy doubles and speed surpasses human-level in most cases, but accuracy drops in scenes with complex, dense motion and dependence on conventional algorithms preserves their limitations.

Source: Robots with human-inspired eyes have better visio

Subtitle: Their reaction times can even surpass their makers’

Dateline: 2月 12, 2026 05:49 上午