Japan's TIER IV open-sources multiple end-to-end driving models, from cameras-only to LiDAR and maps

The backstory: why self-driving split into two paths

A typical autonomous-driving stack works like an assembly line: cameras and lidar first identify lanes, vehicles, and pedestrians, then a planner decides how to steer and accelerate. That makes each step easier to inspect. The weakness is that the modules can talk past each other, and information lost in the middle is hard to recover later.

The other path is end-to-end driving: one AI model looks at the sensor data and directly outputs a trajectory. It may be more flexible, but it is also more of a black box, and proving it is reliable on real roads is hard. Autoware, a widely used open-source driving stack, has historically been closer to the modular approach. TIER IV is one of its main backers.

What was released

On November 11, TIER IV announced that it has developed and open-sourced multiple end-to-end driving models that can be chosen according to map information and sensor setup, from cameras alone to combinations that add lidar, radar, and high-definition maps. The company is also leading Autoware 2.0, which is meant to combine these AI models with conventional rule-based methods so automakers and partners can build their own stacks.

Several models were named. Alpamayo 2 Edge, from TIER IV's collaboration with NVIDIA, takes only camera images and uses reasoning to generate a trajectory. It has about 2 billion parameters, far fewer than the roughly 34 billion in Alpamayo 2 Super, which TIER IV says makes on-vehicle use more realistic. The company said it has finished integrating the model with Autoware and is optimizing it for real vehicles. AutoE2E, led by the Autoware Foundation and TIER IV, can run on hardware beyond NVIDIA. It has about 79.9 million parameters and a fast-slow design: a reactive model for real-time decisions, a world model that predicts how the scene will change, and a reasoning model that helps with judgment. The reactive model has been released and integrated with Autoware; the other two parts are still in development, and the companies are considering robotaxi commercialization. TIER IV's own METEOR is lighter still, at about 16 million parameters, aimed mainly at privately owned cars, does not rely on high-definition maps, can use one to eight cameras, and can also take lidar and radar, with support for chips including NVIDIA and Renesas. Other models, including OnePlanner and LatentDrive, were also announced; Autoware integration and in-vehicle validation for those are still continuing.

Those figures and timelines come from TIER IV's announcement. Wiring a reactive model into an open-source stack is not the same as already running driverless in a city. The company still describes Level 4, meaning fully driverless operation in defined conditions, as a development goal, and tests in real vehicles are ongoing.

What it could mean for you

If these open models later hold up in real vehicles, automakers might not have to build an end-to-end system from scratch. They could choose a lower-cost, camera-only driver-assistance setup or a higher-spec system with more sensors. For passengers, the nearer prospect is still gradually more capable assistance, not streets suddenly full of robotaxis.

What is confirmed now is that a widely used open-source driving stack has, for the first time, put out several end-to-end models that can switch among sensor and map configurations, and that some of them are already plugged into the existing framework. Whether they are both safe enough and useful enough in real traffic will depend on the vehicle tests still to come.

Original source published: TIER IV, Nov 11, 2026, 13:00 Beijing time (UTC+8)