Wayve on Wednesday launched Lingo-2, an artificial intelligence (AI)-based visual language action driving model (VLAM). Lingo-2 is the successor of the Lingo-1 AI model and has several new features. Self-driving AI can now provide commentary on its behavior while driving and adjust its behavior based on instructions from passengers. It can also answer queries about the surrounding environment that are not directly related to driving. Lingo-2 is designed to build trustworthy self-driving technology, the artificial intelligence company said.

Showcasing Lingo-2’s capabilities in a demo video on The post on X also includes a video of Lingo-2 driving through central London, with the model generating real-time driving commentary while driving the car.

The AI ​​model combines three different architectures—computer vision, large language models (LLM), and action models—to create a combined VLAM model that can perform a variety of complex tasks in real time. According to the demonstration, Lingo-2 can see what’s happening on the road, make decisions based on them, and communicate those decisions to passengers. Additionally, it can adapt its behavior to any instructions given by passengers and answer non-driving-related queries, such as information about the weather.

Wayve says performing these operations consistently and reliably is an important step in building self-driving technology. “By incorporating descriptions of driving behavior and causal inference into the training of the model, it opens up new possibilities for accelerating natural language learning.

Even in the future, natural language interfaces will allow users to converse with driving models, making it easier for people to understand these systems and build trust. ” on its website.

It’s worth noting that Lingo-2 doesn’t actually drive the vehicle, as it’s just an artificial intelligence model that’s not integrated with the hardware to control the vehicle. It was trained and tested in Wayve’s in-house closed-loop simulation called Ghost Gym.

As a closed-loop simulation, the company can test the realistic reactions of other vehicles and pedestrians based on the behavior of the controlled vehicle. Next, the AI ​​company plans to begin limited testing of the AI ​​model in real-world settings to analyze its decision-making in more unpredictable situations.