New Camera Tech Lets Robots See in Near-Dark at Blazing Speed
Imagine your car's night vision suddenly 10,000 times sharper.
Scientists have found a way for machines to "see" by processing individual photons as they arrive, without assembling full image frames first. Their approach, called probabilistic events, uses Bayesian inference to track when a scene's brightness last changed, generating real-time signals for motion, activity, and uncertainty. It can estimate the pose of a running person in extreme low light—around 0.05 lux—without retraining existing vision models, and it handles over 50,000 quanta frames per second on commodity GPU hardware, delivering kilohertz-scale outputs up to four orders of magnitude faster than state-of-the-art quanta reconstruction baselines, even for megapixel sensors. By replacing frame reconstruction with direct probabilistic inference, this work aims to bring photon-counting quanta sensing into robotic vision.
- Works in near-total darkness: sees at 0.05 lux, darker than a moonlit night.
- Handles over 50,000 frames per second on a standard GPU, making real-time perception possible.
- No need to retrain existing AI models, so it could speed up adoption in self-driving cars and robotics.
Why It Matters
Safer self-driving cars and rescue drones that can see clearly in the dark, at speeds that prevent crashes.