AI Can’t Do Physics Well – And That’s a Roadblock to Autonomy
AI Can’t Do Physics Well – And That’s a Roadblock to Autonomy
https://hai.stanford.edu/news/ai-cant-do-physics-well-and-thats-a-roadblock-to-autonomy
Publish Date: 2026-01-26 17:23:00
Source Domain: hai.stanford.edu
- The article highlights the struggles of AI in understanding the physical world, particularly in estimating the speed, size, and acceleration of objects, a challenge that impedes progress in robotics and autonomous vehicles.
- QuantiPhy, a new test developed by researchers from Stanford’s STAI and SVL labs, evaluates AI’s ability to comprehend and perform quantitative reasoning in physical interactions.
- QuantiPhy aims to measure AI’s capacity for physical comprehension and guide improvements in models that understand visual, linguistic, and textual inputs.
- The study reveals that current AI models primarily rely on pretrained world knowledge rather than quantitative reasoning from sensory inputs, suggesting a shift from guesswork to more accurate physical understanding.
- The researchers used video data from both the internet and lab-recorded experiments to develop QuantiPhy, finding that an unprompted, end-to-end learning approach works best.
- The team discovered that vision-language models depend too heavily on memorized facts instead of visual inputs, resulting in outputs that lack grounding in quantitative measurements.
- Improved physical reasoning through QuantiPhy has potential applications in healthcare robotics, automotive safety, and domestic robot capability for better environmental interaction.
- Future goals for QuantiPhy include refining spatial calculations in three dimensions and enhancing vision-language models to handle complex dynamics, deformable objects, and multi-camera perspectives.
