Skip to content
Thursday 19 November 2026 · Automotive technology news for engineers

TIER IV Open-Sources End-To-End Autonomous Driving AI Models

TIER IV has released open-source end-to-end autonomous driving AI models for Autoware, including AutoE2E, METEOR, OnePlanner and LatentDrive, alongside an Autoware integration of NVIDIA’s camera-only Alpamayo 2 Edge, covering sensor sets from cameras alone to LiDAR, radar and HD maps.

ATN Editorial Team Source: tier4.co.jp 9 min read
OnePlanner end-to-end autonomous driving model planning trajectories through urban intersections over LiDAR point clouds and a vector map
Credit: Autoware Foundation / TIER IV.

Summary: TIER IV, which pioneered the Autoware open-source software for autonomous driving, has developed and released multiple end-to-end autonomous driving AI models that support different map and sensor configurations, from camera-only set-ups to those using LiDAR, radar and high-definition (HD) maps. They comprise NVIDIA Alpamayo 2 Edge, integrated with Autoware through TIER IV’s collaboration with NVIDIA, plus AutoE2E, METEOR, OnePlanner and LatentDrive, with source code on GitHub and models on Hugging Face. Integration with Autoware is complete for Alpamayo 2 Edge, OnePlanner and the AutoE2E Reactive Model, while LatentDrive integration and in-vehicle validation are under way.

Key engineering takeaway: The end-to-end autonomous driving models span a wide range of sizes. Alpamayo 2 Edge is a camera-only vision-language-action (VLA) model that uses reasoning to generate trajectories, with 2 billion parameters against 34 billion for Alpamayo 2 Super; AutoE2E has 79.9 million parameters and a Fast-Slow architecture combining Reactive, World and Reasoning models; and METEOR, a 16-million-parameter convolutional neural network, runs from perception to trajectory generation in a single network on a bird’s-eye-view (BEV) representation, with one to eight cameras and no HD maps. OnePlanner feeds BEVFusion LiDAR features directly into motion planning and generates trajectories for the next six seconds, and LatentDrive plans four seconds ahead from V-JEPA latent features of camera video, without object detection and tracking stages.

Why it matters: TIER IV says automakers and component suppliers need to internalise AI technology according to their applications and performance requirements, without depending on a particular AI model or semiconductor platform. Releasing several end-to-end autonomous driving models of different sizes and sensor needs, with METEOR running on NVIDIA and Renesas platforms and AutoE2E on hardware beyond NVIDIA, lets developers select a model to suit their requirements and optimise it for their own vehicles through Autoware.

Multiple open-source AI models are available with Autoware

TIER IV, the pioneering force behind Autoware, open-source software for autonomous driving, today announced the development and open-source release of multiple end-to-end (E2E) autonomous driving AI models that support different system requirements, including map information and sensor configurations. The models offer flexible options ranging from camera-only configurations to those incorporating LiDAR, radar and high-definition (HD) maps, enabling developers to select AI models suited to their performance requirements. Through Autoware, TIER IV provides a development environment for optimizing these models and integrating them into vehicle systems, helping automakers and technology partners develop and deploy their own autonomous driving stack.

TIER IV is leading the development of Autoware 2.0, the next-generation architecture of Autoware. Autoware 2.0 is a scalable and adaptable architecture that flexibly combines AI models with conventional rule-based technologies. By supporting different AI models and sensor configurations, it enables autonomous driving systems to be built for a range of applications and system requirements. The E2E autonomous driving AI models announced today are concrete technological outcomes designed to leverage this architecture.

NVIDIA Alpamayo 2 Edge And AutoE2E Models

As part of its collaboration with NVIDIA announced in March 2026, TIER IV has been evaluating the performance of NVIDIA Alpamayo, a vision-language-action (VLA) foundation model, and integrating it with Autoware. The latest model, Alpamayo 2 Edge, is an E2E autonomous driving AI model that takes only camera images as input and uses reasoning to generate driving trajectories. With 2 billion parameters, compared with 34 billion in Alpamayo 2 Super, it is substantially smaller and offers promise for practical deployment in vehicles. TIER IV is currently optimizing the model and adapting it to real vehicles, while having completed its integration with Autoware.

AutoE2E, whose development is led by the Autoware Foundation (AWF) and TIER IV, is an E2E autonomous driving AI model capable of running on computing platforms beyond NVIDIA hardware. It adopts a Fast–Slow architecture combining a Reactive Model for real-time driving decisions, a World Model for predicting future changes in the surrounding environment, and a Reasoning Model that supports driving decisions. It can accommodate configurations ranging from cameras alone to those incorporating LiDAR and radar sensor data. With 79.9 million parameters, it is lighter than Alpamayo 2 Edge. The Reactive Model has been released, while development of the World Model and Reasoning Model continues, alongside consideration of commercialization for robotaxi applications. Integration of the Reactive Model with Autoware has been achieved.

METEOR, OnePlanner And LatentDrive End-To-End Autonomous Driving Models

Developed independently by TIER IV, METEOR is a lightweight, multitask E2E autonomous driving AI model primarily intended for privately owned passenger vehicles. Built on a convolutional neural network (CNN) architecture, it uses a bird’s-eye-view (BEV) representation to perform everything from perception of the surrounding environment to driving trajectory generation within a single neural network, without relying on high-definition maps. It supports configurations using images from one to eight cameras and can also incorporate LiDAR and radar sensor data. With 16 million parameters, it is even lighter than AutoE2E and supports operation across different semiconductor platforms, including NVIDIA and Renesas. TIER IV also uses CoMET, the automated labeling platform within Co-MLOps, together with agentic AI to automate data labeling and model training, evaluation, and improvement, eliminating the need for manual coding and labeling. Automakers and component suppliers can optimize the open-source model to meet their own vehicle specifications and performance requirements.

OnePlanner is an E2E autonomous driving AI model that integrates perception of the surrounding environment with motion planning. It currently uses BEVFusion, which generates a BEV representation from LiDAR point-cloud data, with a planned transition to PTv3. Its flexible architecture allows the perception model to be replaced while directly using features extracted during perception for motion planning. Conventional autonomous driving systems convert perception results into information such as object positions and velocities before passing them to the planner. OnePlanner reduces these intermediate conversions, enabling richer use of information about the surrounding environment. It also incorporates HD maps, route information, traffic-signal states, and vehicle motion states to generate driving trajectories for the next six seconds. Integration with Autoware has been achieved.

LatentDrive is an E2E autonomous driving AI model that uses a world model to integrate predictions of the surrounding environment with motion planning. Using V-JEPA, a self-supervised learning approach, it represents features of the surrounding environment extracted from camera video in a latent space and uses these features to generate vehicle trajectories. Its distinguishing feature is that it directly leverages information from video for motion planning without intermediate processes such as object detection and tracking. It does not require HD maps and generates trajectories for the next four seconds using camera video, route information, and vehicle motion states. It supports configurations with either one or three cameras, and work is underway on integration with Autoware and validation in an in-vehicle environment.

An Open-Source Autonomous Driving Ecosystem Around Autoware

“Advances in E2E autonomous driving are fundamentally changing how autonomous driving systems are developed.” said Shinpei Kato, founder and CEO of TIER IV. “For automakers and component suppliers to incorporate AI technology into their own products, however, it is essential that they be able to internalize technologies according to their applications and performance requirements, without depending on a particular AI model or semiconductor platform. By releasing multiple AI models developed under the leadership of TIER IV and the Autoware Foundation, alongside the results of our collaboration with NVIDIA, we have significantly expanded those choices. We will continue to advance the open-source ecosystem centered on Autoware and work with partners around the world to democratize autonomous driving.”

“Advancing autonomous driving requires AI that can reason through complex situations and run efficiently in the car.” said Ali Kani, vice president of automotive at NVIDIA. “TIER IV’s integration of NVIDIA Alpamayo 2 Edge with Autoware gives developers an open foundation to bring that intelligence from research into their own vehicles.”

“The advancement of autonomous driving technology depends on an ecosystem in which diverse AI models can be developed and evaluated openly, then applied according to their respective use cases.” said Muhammad Zain Khawaja, President of the Autoware Foundation. “The AI models released today are an important demonstration of how Autoware is evolving into a flexible development platform that is not tied to any particular technology or hardware. We welcome TIER IV’s continued technology development and its collaboration with partners including NVIDIA. The Autoware Foundation will continue working with the global open-source community to make advanced autonomous driving technology accessible to everyone.”

TIER IV will continue improving the performance of the AI models released today and optimizing them for in-vehicle environments, while promoting the development and commercialization of E2E autonomous driving technology by automakers and partners through Autoware. TIER IV will also strengthen collaboration with the global open-source community and work to expand the adoption of autonomous driving systems across diverse vehicle types and applications.

End-To-End Autonomous Driving: Frequently Asked Questions

What is end-to-end autonomous driving?

End-to-end autonomous driving is an approach in which a neural network maps sensor input directly to a planned driving trajectory, rather than passing results between separately engineered perception, prediction and planning modules. TIER IV’s release describes the contrast: conventional autonomous driving systems convert perception results into information such as object positions and velocities before passing them to the planner, whereas its OnePlanner model reduces these intermediate conversions.

Which end-to-end autonomous driving models has TIER IV released for Autoware?

The release covers NVIDIA Alpamayo 2 Edge (2 billion parameters, camera-only, integrated with Autoware), AutoE2E (79.9 million parameters, led by the Autoware Foundation and TIER IV, Reactive Model released), METEOR (16 million parameters, developed independently by TIER IV, primarily for privately owned passenger vehicles), OnePlanner (LiDAR-based perception integrated with motion planning, integrated with Autoware) and LatentDrive (world-model-based planning from camera video, with Autoware integration under way).

What is a vision-language-action (VLA) model?

A vision-language-action model is a neural network that combines visual perception with the reasoning capability of a large language model and outputs actions, which in a vehicle means driving trajectories. In the release, NVIDIA Alpamayo 2 Edge takes only camera images as input and uses reasoning to generate driving trajectories, with 2 billion parameters compared with 34 billion in Alpamayo 2 Super.

What is Autoware 2.0?

According to the release, Autoware 2.0 is the next-generation architecture of Autoware, the open-source autonomous driving software, and is a scalable and adaptable architecture that flexibly combines AI models with conventional rule-based technologies. TIER IV is leading its development, and the end-to-end models are designed to leverage this architecture.

Do the models need HD maps?

Not all of them. METEOR and LatentDrive do not require high-definition maps, and Alpamayo 2 Edge takes only camera images as input. OnePlanner incorporates HD maps, route information, traffic-signal states and vehicle motion states to generate driving trajectories for the next six seconds.

Further Reading

For more autonomous driving news, click here.

Source

The latest technology and engineering news direct to your inbox.

ATN publishes first-party announcements within seven days of release. The summary at the top is written by ATN for engineers; the text below it is the company's own.