HaoMo Zhixing: The large autonomous driving cognitive model has been officially upgraded to DriveGPT

According to news on February 17, Haomo Zhixing announced that its large autonomous driving cognitive model has been officially upgraded to DriveGPT, and will announce its latest important progress at HAOMOAI DAY in April 2023.

As ChatGPT becomes popular across the entire network, the Transformer large model and "Human Feedback Reinforcement Learning (RLHF)" technology it uses has once again attracted industry attention. In the field of autonomous driving, Feimo was the first in China to introduce the Transformer large model into the data intelligence system MANA. At the HAOMO AI DAY in January 2023, Gu Weihao, CEO of Haomo Zhixing, introduced that the large self-supervised cognitive model for human driving launched by Haomo has drawn on the implementation ideas of ChatGPT and adopted RLHF (Human Feedback Reinforcement Learning) technology. Introducing real human driver takeover data to continuously optimize the autonomous driving cognitive decision-making model.

Haimo’s cognitive decision-making algorithm has gone through three stages of evolution. The first stage is to introduce end-to-end imitation learning of individual scenarios to directly fit human driving behavior. The second stage is to use large cognitive models to introduce massive amounts of normal human driving data, and achieve controllable and explainable cognitive decision-making through prompts. The third stage, which is the current stage, introduces real takeover data and begins to use the RLHF algorithm in a large model to learn from human-driving takeover data. To this end, Hao Mo built a Pairwise ranking model including "old strategy, takeover strategy, and manual label strategy". Based on this ranking model, Haimo built a reward model for autonomous driving decisions to make optimal decisions under various planning scenarios. Through this large model, in recognized difficult scenes such as U-turns and roundabouts, the scene pass rate is increased by more than 30%.

#On February 17, Haimo Zhixing officially upgraded its large-scale human driving self-supervision cognitive model to "DriveGPT", which is also the world's first large-scale autonomous driving cognitive model. At present, Haimo DriveGPT has completed model construction and first-stage data running, and the parameter scale can benchmark the level of GPT-2. Next, DriveGPT will continue to introduce large-scale real takeover data and continuously improve the evaluation effect through reinforcement learning of human driving data feedback. DriveGPT will also be used as a cloud evaluation model to evaluate the driving effect of small car-side models. The latest important progress of Haomo DriveGPT will be announced at the 8th HAOMO AI DAY in April 2023.

The above is the detailed content of HaoMo Zhixing: The large autonomous driving cognitive model has been officially upgraded to DriveGPT. For more information, please follow other related articles on the PHP Chinese website!