The Pulse
TIER IV releases driving AI models for different sensor setups
TIER IV has released multiple open-source autonomous driving AI models for use with its Autoware software. The models differ in their sensor and map requirements, and some integration and development work remains underway.
AI.info Team · Written with AI · fact-checked ·
TIER IV has released multiple open-source autonomous driving AI models designed for different sensor and map configurations, the company announced in a November 11 release. The models range from systems using cameras alone to those that also incorporate LiDAR, radar or high-definition maps.
The release accompanies work on Autoware 2.0, the next-generation architecture of the open-source autonomous driving software. TIER IV says Autoware provides an environment in which developers can optimize models and integrate them into vehicle systems.
A camera-only NVIDIA model
Among the models is NVIDIA Alpamayo 2 Edge, which takes camera images as its only input and uses reasoning to generate driving trajectories. The model has 2 billion parameters, compared with 34 billion in Alpamayo 2 Super.
TIER IV has completed the model's integration with Autoware and is optimizing it and adapting it to real vehicles. The work is part of a collaboration with NVIDIA announced in March 2026, under which TIER IV has been evaluating and integrating the Alpamayo vision-language-action foundation model.
Models for different hardware
The Autoware Foundation and TIER IV lead development of AutoE2E, which can run on computing platforms beyond NVIDIA hardware. Its architecture combines a Reactive Model for real-time driving decisions, a World Model for predicting changes in the surrounding environment and a Reasoning Model to support driving decisions.
At 79.9 million parameters, AutoE2E is smaller than Alpamayo 2 Edge and can use cameras alone or add LiDAR and radar data. The Reactive Model has been released and integrated with Autoware; the other two components remain in development, while commercialization for robotaxi applications is under consideration.
TIER IV independently developed METEOR, a 16 million-parameter model intended primarily for privately owned passenger vehicles. It can use images from one to eight cameras, incorporate LiDAR and radar, and operate without high-definition maps.
The company says METEOR uses a single neural network for tasks from perceiving the environment to generating driving trajectories. It also supports different semiconductor platforms, including NVIDIA and Renesas, while automakers and component suppliers can optimize the model for their vehicle specifications and performance requirements.
Two approaches to motion planning
OnePlanner combines perception with motion planning and currently uses BEVFusion to produce a bird's-eye-view representation from LiDAR point-cloud data. A transition to PTv3 is planned, and the model's architecture allows its perception component to be replaced.
According to TIER IV, OnePlanner uses features extracted during perception directly for planning, reducing intermediate conversions of perception results. It also uses high-definition maps, route information, traffic-signal states and vehicle motion states to generate trajectories for the next six seconds; integration with Autoware is complete.
LatentDrive uses a world model to combine predictions about the surrounding environment with motion planning. It generates trajectories for the next four seconds from camera video, route information and vehicle motion states, without requiring high-definition maps.
The model supports one or three cameras. Its integration with Autoware and validation in an in-vehicle environment are still underway.