May Mobility on Wednesday launched its fifth-generation autonomy system, an on-vehicle architecture that fuses deep learning with a predictive world model and the company’s reasoning engine, built to scale across more than 525,000 commercial rides already logged. The Ann Arbor, Michigan-based autonomous vehicle company said the system runs hundreds of “what if” simulations every 200 milliseconds, evaluating possible futures up to 10 seconds ahead before selecting a driving action. The architecture is rolling out across May Mobility’s existing fleet and will support its upcoming deployment on the Uber platform in Arlington, Texas. Full details on the system are published on the company’s technology page.
Highlights
- Fifth-generation autonomy system integrates deep learning, a predictive world model, and May Mobility’s reasoning engine on a single on-vehicle architecture
- More than 525,000 commercial rides and 1.1 million autonomous miles completed to date, with driverless deployments in three U.S. states
- World model evaluates hundreds of “what if” simulations every 200 milliseconds, projecting up to 10 seconds into the future
- New architecture supports the upcoming May Mobility deployment on the Uber platform in Arlington, Texas
A Hybrid Approach to AV Architecture
May Mobility’s new system departs from both modular AV stacks and end-to-end neural network models. The company describes the architecture as a combination of deep learning with reasoning, allowing the vehicle to draw on training data while accounting for context the data may not represent.
The aim, according to May Mobility, is to generalize across novel situations, new geographies, and complex driving environments without the data and compute footprint that end-to-end systems typically require.
“Driving by memorization is bad—humans don’t need to see a billion miles of road to drive safely. The brain instantly builds a mental model of the world and then reasons through it. Our new system approaches driving the same way, and it dramatically changes how autonomy can safely scale,” said Dr. Edwin Olson, CEO of May Mobility.
World Model and Reasoning Engine
Two fused components run together on the vehicle.
The world model processes the environment around the vehicle through what the company describes as a distillation of physics, rules of the road, and driving culture. Applied repeatedly, it produces hundreds of forward simulations every 200 milliseconds, each one representing a possible future. The model predicts how every road user’s behavior will affect every other, simulating up to 10 seconds ahead per scenario.
The reasoning and planning engine uses May Mobility’s multi-policy approach, in which multiple driving strategies compete to control the vehicle based on how each performs against the simulated futures generated by the world model. The system simulates outcomes of both deep-learned and rule-based strategies, rejecting any that fail safety parameters. Vehicle control, the company says, is always traceable back to its source — a contrast with end-to-end models, where decision provenance is opaque.
Smaller Models, Lower-Cost Hardware
May Mobility positions the fifth-generation system as a path to lower-cost AV hardware. The company argues that conventional AV stacks built to memorize driving situations require collecting and training on massive datasets, producing models large enough to demand custom compute.
Because May Mobility’s models are built around understanding how the world works rather than memorizing examples of it, the company says they can be smaller, enabling the use of less expensive hardware to handle long-tail edge cases. The company frames this as a challenge to the industry assumption that massive datasets and custom silicon are prerequisites for full autonomy.
Deployment Path
May Mobility has begun rolling the update across its current fleet and says the new architecture will enable additional driverless deployments in the near future. Ride-hail networks are positioned to be among the first to encounter the system, with the Uber deployment in Arlington, Texas, named explicitly.
The company has previously announced partnerships with Toyota Motor Corporation, NTT, Lyft, Uber, and Grab, and operates Autonomy as a Service across the U.S. and Japan. Its first U.S. driver-out deployments have run in Sun City, Arizona; Ann Arbor, Michigan; and Peachtree Corners, Georgia.
Sign up for our weekly emails to stay up to date on the latest news!
