TIER IV and TUM establish Autoware research centre

TIER IV and the Technical University of Munich have established a research centre to develop safety supervision, verification methods and repeatable deployment processes for autonomous driving software. Announced on 15 October, the multi-year collaboration will use the open-source Autoware platform as its foundation, bringing together three university research groups specialising in vehicle engineering and autonomous systems.

The TUM Autoware Research Center by TIER IV will receive strategic research funding from the Japanese autonomous driving technology company. The participating groups are the Chair of Automotive Technology, the Professorship of Cyber Physical Systems and the Professorship of Autonomous Vehicle Systems. Their work will cover safety, evaluation and scalability, with an emphasis on making autonomous driving software more consistent across development and deployment environments.

Autoware provides an open-source software framework for developing autonomous vehicle systems, connecting functions such as localisation, perception, planning and vehicle control. These functions interact with sensors, vehicle networks, processors and actuators, requiring software to operate within the timing, computational and safety constraints of the physical platform.

Vehicle localisation combines information from available sensors and positioning systems to estimate the vehicle’s location and orientation. Perception software interprets surrounding objects and road conditions, while planning algorithms determine an intended trajectory. The control system must then translate that trajectory into commands compatible with the vehicle’s steering, braking and propulsion capabilities.

Errors or uncertainty can propagate between these functions. A perception system may incorrectly identify an object or estimate its movement, while a planning system may propose a trajectory that becomes unsuitable as conditions change. The resulting behaviour depends on the information available to the software and its ability to identify and respond to constraints before the physical vehicle reaches an unsafe state.

The first research area will investigate a hybrid safety architecture combining learning-based driving models with rule-based supervision. Researchers intend to develop a verifiable safety guardian that monitors vehicle behaviour and intervenes when defined safety constraints are approached or violated. The guardian is a proposed research outcome, not a completed system that has already passed an independent certification process.

Learning-based models can represent complex relationships observed during training, but their behaviour may be difficult to characterise exhaustively across every possible operating condition. A rule-based supervisor can provide a separate mechanism for checking whether certain actions remain within defined limits. The effectiveness of that separation depends on the accuracy of the information available to the supervisor and whether the system can take an appropriate action within the required response time.

A supervisory system may need to evaluate vehicle speed, available braking distance, surrounding traffic and the limits of the current driving environment. If an intended manoeuvre breaches a defined constraint, the architecture must determine how to change or interrupt the action. That intervention depends on the vehicle’s control interface and physical capabilities; software supervision cannot compensate for an actuator failure or a situation in which a safe response is no longer physically possible.

The second workstream concerns benchmarking and verification across both component and complete-system levels. TIER IV and TUM plan to connect simulation, offline analysis and real-world operation within a repeatable assessment framework. This will allow changes to individual software functions to be evaluated alongside their effects on the behaviour of the wider vehicle system.

Simulation provides controlled conditions for examining repeatable scenarios, including events that would be difficult or unsafe to reproduce deliberately on public roads. Development teams can vary vehicle behaviour, sensor information and the actions of other road users to assess how software responds. Such testing depends on the fidelity of the simulated environment, however, and cannot establish that every physical interaction has been represented accurately.

Offline analysis of recorded data provides another route to comparing software implementations without repeating every physical test. The same inputs can be processed through different software versions to identify changes in their outputs, while real-world trials provide evidence about vehicle dynamics, sensor behaviour and environmental conditions. Combining the methods can improve the breadth of evaluation, but none provides comprehensive evidence of safety independently of the others.

The research will build on TUM’s EDGAR autonomous vehicle and its experience with CommonRoad, a framework used in research into vehicle motion planning and evaluation. EDGAR already operates with Autoware and provides a practical platform for testing how open-source driving software performs in physical environments. The new centre therefore builds on an existing research vehicle rather than announcing the first deployment of Autoware at the university.

During IAA Mobility 2025, EDGAR provided public demonstration rides between Messe München and the city centre, covering different road conditions. TUM has also demonstrated adaptation of the vehicle for autonomous operation in Berlin within approximately one week. The Berlin exercise, conducted with TÜV-Verband, examined a Level 4 assessment methodology under real-world traffic conditions for the German government.

Those demonstrations provide evidence of existing research capabilities but do not establish that deployment to every new city can be completed in the same period. Local road layouts, traffic rules, infrastructure, vehicle calibration and operating conditions can require different engineering work. The new research programme will examine methods intended to make this process more consistent and repeatable.

Scalability is the third research area. The partners plan to establish repeatable procedures for adapting and evaluating Autoware across new environments, with suitable improvements contributed to open-source projects. These activities will involve identifying which configuration and verification steps can be reused and which require changes specific to the vehicle or its operating location.

Open-source development provides a mechanism for sharing software components and improvements, although a common codebase does not guarantee identical operation on different hardware. Processor performance, sensor interfaces, communications timing and actuator characteristics can all affect the behaviour of autonomous driving software. A reusable deployment process must therefore accommodate differences in the underlying electronics and vehicle architecture.

Shinpei Kato, founder and chief executive officer of TIER IV, identified AI, system architecture, verification and deployment as the main technical areas supporting the collaboration. Johannes Betz, Professor of Autonomous Vehicle Systems at TUM, emphasised combining learning-based techniques with conventional rule-based methods to pursue systems that can be assessed for safety and deployed in real environments.

The initiative also provides a base for doctoral training, software development and cooperation with industrial and public transport projects. The partners intend to support demonstrations and benchmark challenges, allowing research methods to be evaluated through practical examples. Specific project budgets, completed safety-supervision results and a timetable for the first new benchmark release have not been disclosed.

The European Commission has separately outlined an approach to software-defined vehicle development that includes open-source software, reference implementations, standardised components and shared validation tools. The research centre is consistent with that wider technical direction, but its announcement does not establish a new European Commission grant or a formal certification role for the university.

The confirmed October development is the establishment of the multi-year research collaboration and its three defined workstreams. The proposed safety guardian, benchmarking framework and repeatable deployment methodology remain subjects of future engineering work. Their progress will depend on published results, practical demonstrations and evidence that the methods can be reproduced across different vehicle platforms and operating conditions.