TIER IV Releases Five Autonomous Driving Models for Autoware
TIER IV unveils five autonomous driving AI models for Autoware, with variants using camera-only, LiDAR, and radar. AutoE2E v1.0 is also released under Apache 2.0 license, though real-world validation is ongoing.
On November 11, 2026, TIER IV announced five autonomous driving AI models for Autoware, its open-source software for autonomous vehicles. The set covers multiple architectures, from camera-only to configurations with LiDAR, radar, and HD maps, to help manufacturers and partners choose based on their sensors and computing power. Simultaneously, the Autoware Foundation released AutoE2E v1.0 under the Apache 2.0 license. The scope is clear, but the path to real deployment remains long.
Une gamme pensée pour des véhicules très différents
The five models are named NVIDIA Alpamayo 2 Edge, AutoE2E, METEOR, OnePlanner, and LatentDrive. All link environment perception to trajectory generation but with different trade-offs regarding sensors, reasoning, and vehicle integration. In other words, TIER IV does not push a single magic recipe but several building blocks adapted to distinct uses.
Alpamayo 2 Edge, for example, has 2 billion parameters compared to 34 billion for Alpamayo 2 Super. It operates solely from camera images and reasoning, and its integration into Autoware is already complete, although optimization and vehicle adaptation continue. NVIDIA and TIER IV present it as a bridge between research and industrialization, which in this sector often remains the toughest part of the menu.
AutoE2E, developed by the Autoware Foundation and TIER IV, relies on 79.9 million parameters and can run beyond NVIDIA platforms. Its fast-slow architecture combines a reactive model for real-time decisions with world and reasoning modules tasked with anticipating the environment. Configurations range from simple camera setups to systems integrating LiDAR and radar.
Des approches techniques qui évitent le copier-coller
METEOR mainly targets passenger vehicles. TIER IV indicates this 16-million-parameter network uses a bird’s-eye view representation to manage perception through to trajectory without relying on HD maps. It accepts images from one to eight cameras, can add LiDAR and radar, and supports NVIDIA and Renesas semiconductor platforms.
OnePlanner follows a different logic. The model combines perception and motion planning, relies on BEVFusion to convert LiDAR point clouds into a bird’s-eye view, then directly transmits useful information to the planner. This flow reduces intermediate conversions into object positions and speeds. The system adds HD maps, route information, traffic light status, and vehicle state to produce trajectories over six seconds. Its integration into Autoware is already complete.
LatentDrive, finally, uses V-JEPA, a self-supervised learning method that learns latent representations from camera videos. The model does not require HD maps and generates trajectories over four seconds from video, route, and vehicle state. It supports one or three cameras, but its integration into Autoware and onboard validation remain ongoing.
AutoE2E v1.0 avance, sans confondre test et conduite réelle
The Autoware Foundation released AutoE2E v1.0 on November 11, 2026, under the Apache 2.0 license, with trained checkpoints available for research and commercial use. The version includes the Reactive Model, which runs at 10 Hz, merges six camera views into a bird’s-eye representation, and plans 6.4 seconds of acceleration and steering. The world and reasoning layers still need development to work with the same planner.
On recorded data, the refined model shows an average displacement error of 0.744 meters over three seconds, according to the foundation, ranking second among eight compared open-weight configurations. However, this result concerns an open-loop evaluation, i.e., on trajectories predicted from recorded video sequences, not on a driverless car in traffic. The project targets uses ranging from hands-free assistance at L2++ level to L4 mobility, but level L4 autonomous driving remains a development goal, not an achieved capability.
Ce que cherche TIER IV derrière cette stratégie
TIER IV’s message is less technological than industrial: to allow manufacturers and suppliers to “internalize technologies according to their applications and performance requirements, without depending on a particular AI model or semiconductor platform,” according to Shinpei Kato. This logic matters because autonomous driving is not only about a smarter model but about its ability to adapt to sensors, embedded computing, and production constraints.
To accelerate METEOR, TIER IV also relies on CoMET, its automated labeling platform integrated with Co-MLOps, combined with agentic AI to automate data labeling, training, evaluation, and model improvement. The company now aims to improve performance, optimize embedded systems, commercialize these building blocks with Autoware, and extend their use to more vehicles. One decisive nuance remains: open source does not yet mean validated deployment, and that is often where schedules start to falter.
Key points
- November 11, 2026: release of five models for Autoware.
- AutoE2E v1.0 released under Apache 2.0 license.
- Alpamayo 2 Edge has 2 billion parameters.
- AutoE2E has 79.9 million parameters.
- The Reactive Model runs at 10 Hz.
- TIER IV cites Shinpei Kato and NVIDIA.