UniAD

An end-to-end autonomous-driving framework that chains perception, prediction, and planning into one network, with everything serving the final plan.

UniAD was released by the Shanghai Artificial Intelligence Laboratory's OpenDriveLab (Hongyang Li's group) and other institutions in December 2022, winning the CVPR 2023 Best Paper Award. Traditional autonomous driving splits perception, prediction, and planning into separate modules or parallel multi-task heads optimized independently, which loses information and compounds errors between modules. UniAD's principle is 'planning-oriented': a single network sequentially performs object tracking, online mapping, motion prediction, occupancy prediction, and trajectory planning, with information passed between modules through shared query vectors, and the whole system optimized toward the final planning outcome. It outperformed prior methods across every metric on the nuScenes dataset, becoming a landmark for end-to-end autonomous driving that is also frequently cited when discussing end-to-end design in embodied AI more broadly. The team released a UniAD 2.0 code version in October 2025.

ExampleUniAD takes in images from multiple surround-view cameras on a car, tracks nearby vehicles, draws lane lines, and predicts their movement over the next few seconds, all within the same network, and directly outputs the ego vehicle's upcoming driving trajectory.

Also called

Unified Autonomous Driving, Planning-oriented Autonomous Driving

Related

Autonomous Driving · End-to-End · Tesla FSD v12 · DriveVLM · Bird’s-Eye View · Occupancy Network

Sources

Planning-oriented Autonomous Driving (arXiv 2212.10156)
OpenDriveLab/UniAD (GitHub)
CVPR 2023 Awards

As of

2025-10