Autonomous driving has been undergoing a major shift. Following a long period defined largely by rule-based robotics, approaches built around data-driven development and AI models are in the spotlight. This article outlines the ideas behind this shift, focusing on Autoware, the open-source software at the center of TIER IV’s autonomous driving platforms, and hybrid architectures that combine traditional robotics-based methods with data-centric AI models.
When people think about the benefits of autonomous driving, two ideas often come to mind. The first is safety. About 1.2 million people worldwide lose their lives each year in road traffic accidents. To put that figure into perspective, it is comparable to about ten large passenger planes crashing every day. If such incidents occurred in aviation, few people would feel comfortable flying. Yet in car-dependent societies, this level of risk is something we accept as part of daily life.
According to the World Health Organization, road traffic injuries are the leading cause of death for children and young adults aged 5 – 29 years. As a large proportion of crashes are attributed to human error, autonomous driving is seen as a promising way to reduce serious road accidents over time.
The second benefit is time. In the United States, people spend about an hour a day driving. Once vehicles can reliably handle that task, that hour becomes usable time. For many people, that shift will change how they structure their day, opening space for work, rest, or simply doing something other than focusing on the road.
Looking ahead over the coming decades, it is difficult to identify many technologies with a comparable potential impact on everyday life. Crucially, autonomous driving is no longer a distant vision. Deployments have already begun.
TIER IV is involved in autonomous driving projects with municipalities across Japan. One example is Shiojiri in Nagano Prefecture, where a Level 4 trial has demonstrated autonomous operation without a driver seated at the wheel. The route includes areas with relatively high traffic volumes, and even in these environments, the system has performed reliably.
Urban autonomous mobility services inevitably involve many challenging scenarios, which means large-scale deployment can take time. By contrast, environments such as factories and other private or restricted areas are often easier to implement, since many edge cases can be eliminated by design. These domains have therefore become important early deployment targets.
Open-source solution
As autonomous driving moves from individual pilots toward broader deployment, scalability comes into focus. Even when a system works well on one vehicle model, adapting it to another can be resource-intensive. A simple change in sensor placement is enough to trigger major reconfiguration, increasing development time and cost.
This challenge is something TIER IV has been conscious of from the outset. Autoware is designed to be vehicle-agnostic, supporting a wide range of vehicle types and sensor configurations, and it is publicly available to developers worldwide. It is one of the most widely adopted autonomous driving platforms globally, with a number of organizations using it in research, development and pilot deployments.
In Autoware, sensor data is processed through perception and localization modules. Planning determines how the vehicle should behave and control modules translate those decisions into steering, acceleration and braking commands. While this modular structure is part of Autoware’s foundations, an industry focus in recent years has been on architectures that apply machine learning across the entire pipeline. Autoware supports these approaches alongside traditional modular frameworks.
A defining characteristic of Autoware is its high degree of customizability. Depending on system requirements – such as operating environment, cost constraints, or development stage – teams can select which functions to include. Autoware’s open-source foundations reduce R&D time and cost compared with building a system from scratch.
From AV 1.0 to AV 2.0
In the autonomous driving field, the industry has been shifting from what is referred to as AV 1.0 to AV 2.0.
- AV 1.0 is characterized by iterative development through testing. Engineers identify issues, apply fixes and repeat this cycle. While effective, this approach relies heavily on manual engineering.
- AV 2.0 introduces a data-driven improvement loop. Data is collected and models are trained, deployed, evaluated and refined through continuous feedback. Learning from real-world data becomes central.
This shift is closely tied to the advancement of AI technologies that handle the entire process from sensor input to vehicle control in a data-driven manner. Within AV 2.0, there are different approaches, including those that achieve everything from perception to planning and control using a single end-to-end AI, and those that combine multiple AI models, such as perception AI and planning AI.
Each approach has strengths and limitations.
The AV 1.0 modular stack, which combines AI components such as object detection with traditional rule-based logic, offers high interpretability and adaptability. It performs highly effectively and reliably in environments like factories or private sites where edge cases are limited. However, manually engineered logic has limitations in handling complex traffic situations and rare scenarios.
End-to-end AI has the potential to achieve high performance because the model directly translates sensor input into driving behavior. On the other hand, it faces the challenge of limited interpretability: It is difficult to understand why the system arrived at a particular decision, making it challenging to isolate and debug specific causes.
The approach of combining perception AI and planning AI sits in the middle of this spectrum. By conducting training across each AI component, it achieves data-driven performance improvements while maintaining a certain degree of interpretability. Furthermore, since functions like perception and planning can be validated independently, evaluation and debugging are much more manageable.
Recognizing these trade-offs, TIER IV is developing a hybrid architecture that integrates traditional robotics-based methods, a single end-to-end AI and a combination of AI models, such as perception AI and planning AI. After verifying the safety of each potential outcome, the system adopts the most appropriate decision.
To further push the boundaries of the system, TIER IV is collaborating with NVIDIA, integrating the vision-language-action model NVIDIA Alpamayo. Using language understanding to make sense of complex and changing environments provides the level of judgment needed for advanced, anticipatory driving.
By combining high-performance AI-based models with robust robotics-based logic, the system can raise the baseline for both performance and safety. This hybrid approach has been demonstrated in tests in Tokyo’s Odaiba district.
In tests conducted last year, the system was trained on about 500 hours of synthetic data and about 40 hours of real-world data. The performance was extremely stable. As the volume of data expands, the system's performance continues to improve.
Wrap-up
Autoware provides a collaborative foundation for researchers, startups, automakers and operators to develop systems tailored to specific needs. Because of its diverse ecosystem, the platform improves daily. Building on this foundation, TIER IV’s hybrid architecture, combining robotics with data-centric AI, is an effective way to solve real-world deployment challenges.
Momentum in autonomous driving is growing as research advances and large datasets become more accessible. TIER IV’s open-source approach ensures this progress is shared across the community, strengthening the entire Autoware ecosystem. If you’re interested in learning more, join the conversation on GitHub and show your support by giving the repository a star.
This article is based on a presentation by Yukihiro Saito, head of TIER IV’s Advanced Technology Department.
TIER IV is always on the lookout for passionate individuals to join our journey. If you share our vision of making autonomous driving accessible to all, get in touch.
We’re currently hiring for the following related positions:
Visit our careers page to view all job openings.
If you’re uncertain about which roles align best with your experience, or if the current job openings don’t quite match your preferences, register your interest here. We’ll get in touch if a role that matches your experience becomes available, and schedule an informal interview.
Inquiries
- Media: pr@tier4.jp
- Business: sales@tier4.jp
Social Media
X (Japan/Global) | LinkedIn | Facebook | Instagram | YouTube
More