How can generative AI support software engineers working on robotic systems without compromising reliability, traceability or human oversight?
ASTIR is investigating this question in three different application domains: industrial manufacturing, mobile service robotics and unmanned aerial vehicles for agriculture. Together, these use cases provide realistic environments for developing and evaluating the project’s AI-augmented software engineering methods.
Although the robots, software stacks and operational environments differ considerably, all three domains face a common challenge: robotics software is becoming increasingly complex, while engineering teams must continue to ensure that systems remain reliable, maintainable and understandable.
ASTIR therefore uses the three pilots not simply as demonstrations, but as complementary environments in which its technologies can be assessed against practical engineering tasks. The project’s Robotics Software Engineering Benchmark is correspondingly structured around three pilot-based tracks covering manufacturing, service robotics and UAVs.
Manufacturing Robotics: KUKA
Industrial production systems combine robots, PLC software, mechanical components, sensors and other cyber-physical systems into tightly interconnected assembly lines. Changes to one part of such a system can have consequences elsewhere in the engineering process.
The ASTIR manufacturing use case, led by KUKA Assembly & Test, focuses on software engineering activities associated with the development and validation of these systems. KUKA’s pilot addresses three main areas: PLC code validation, sequence analysis and runtime monitoring.
For example, ASTIR will investigate whether AI-supported tools can help engineers identify inconsistencies when mechatronic components are added to an existing machine configuration. This includes checking PLC logic against engineering information such as device lists and cycle-time diagrams, identifying parameter or naming mismatches, and detecting deviations from expected execution sequences.
Another scenario examines material flow within a simulated robot cell. Movement traces from elements such as the robot gripper, workpiece and pallet are analysed and compared with the expected sequence and timing defined in the engineering specification.
The underlying challenge is therefore not simply generating code. AI-supported engineering results must remain traceable to the underlying technical evidence and understandable to the engineers responsible for the system.
Mobile Service Robotics: PAL Robotics
Service robots face a different engineering environment. They operate in dynamic spaces such as research laboratories, logistics environments and other settings where robots may interact with people and changing surroundings.
PAL Robotics develops a substantial ROS 2-based software ecosystem across mobile manipulators, humanoid research robots and intralogistics platforms. Its developers must continuously integrate new functionality while maintaining many software packages and adapting to rapidly evolving open-source dependencies.
The ASTIR pilot focuses on recurring software engineering tasks where this complexity creates considerable manual effort. The planned evaluation scenarios include AI-assisted migration of ROS 1 packages to ROS 2, generation of integration tests, software vulnerability detection, documentation verification and generation of demonstration applications.
ROS migration illustrates the problem well. Moving an existing package to ROS 2 can require API changes, dependency updates, modifications to launch and message definitions, new tests and updated documentation. ASTIR will investigate whether AI-supported engineering can assist with these transformations while preserving functionality and producing changes that developers can review and validate.
Here, the challenge is scalability: how can AI assist developers across a large and evolving robotics software ecosystem while maintaining software quality, interoperability and developer confidence?
UAVs for Agriculture: UNPARALLEL
The third use case takes ASTIR from factories and indoor mobile robots into agricultural environments.
UNPARALLEL Innovation leads the UAV pilot, which examines AI-augmented software engineering for unmanned aerial vehicles used in precision agriculture. UAVs equipped with multispectral and thermal sensors can provide detailed information about crops, soil and environmental conditions, supporting applications such as crop monitoring and agricultural decision-making.
Developing these systems introduces another combination of constraints. UAV software must interact closely with specialised hardware, operate with limited onboard computing resources, process substantial sensor data and comply with aviation and operational requirements.
ASTIR is investigating support for activities ranging from requirements and software generation to mission planning and dependability assurance. The pilot includes work on interpreting agricultural data, translating mission objectives into symbolic plans and executable flight behaviour, and incorporating regulatory constraints into the engineering workflow.
The UAV case therefore tests whether AI-supported engineering methods can remain useful when software decisions are closely connected to physical environments, operational constraints and regulatory requirements.
Three Domains, Shared Questions
The three ASTIR use cases differ deliberately. KUKA works with PLC-controlled industrial production systems; PAL Robotics works with distributed ROS 2 software for mobile robots; and UNPARALLEL works with UAV software, sensor data and mission planning.
This diversity allows ASTIR to investigate whether its methods can support more than one particular robot or software stack.
At the same time, several engineering questions are shared across all three pilots:
-
Can AI reduce repetitive software engineering effort while maintaining the quality of the resulting artefacts?
-
Can generated or modified software be traced back to requirements and other engineering evidence?
-
Can developers understand why an AI-supported tool produced a particular result?
-
How should AI-generated outputs be tested before they influence physical robotic systems?
-
Where should engineers remain explicitly responsible for reviewing and approving AI-supported decisions?
For this reason, ASTIR combines technical performance measures with qualitative and expert assessment. The project evaluation framework is designed to examine not only performance and efficiency, but also reliability, usability and developer confidence across the three use cases.
As ASTIR moves into its integration and evaluation phase, these pilots will provide the practical environments needed to determine where generative AI can offer meaningful support to robotics engineers—and where further development, verification or human judgement remains necessary.
