Mountain View, California - Waymo, Google's self-driving spin-out, has discovered that the principles governing the performance of autonomous vehicles (AVs) are similar to those of large language models (LLMs).
Research indicates that increasing training data and computational resources directly enhances AV performance. This finding suggests a power-law relationship, where improvements in performance correlate with the scaling of training compute and dataset sizes.
Waymo's research highlights key distinctions between AVs and LLMs. While LLMs often benefit from larger model sizes, AVs may achieve optimal performance with relatively smaller models, provided they are trained on significantly more data.
This insight has important implications for data collection strategies and model size selection in AV development. Smaller model sizes in AVs can lead to lower latency, improving onboard system performance through scaled training dataset size and compute.
Waymo currently has data covering 500,000 hours of driving and uses its 'Carcraft' virtual world for driving simulations. The company operates in several cities, with plans to expand to ten cities this year.
Waymo believes that enriching the quality and size of data and models will lead to better AV performance. This conclusion provides developers with a clear path to improve AV capabilities.