You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
I want to build and test a practical human-override system for Physical AI.
By Physical AI, I mean AI systems that are able to act through robots, drones, vehicles, machines, edge computers, or other physical systems. My concern is quite simple. If these systems become more autonomous, we should not assume that sending a normal software command will always be enough to stop them.
The idea is to build an emergency-control layer that is separate from the AI system itself. If something goes wrong, a human should still be able to limit what the system can do, cut its communication, isolate it, move it into a safe mode, or stop it completely.
I would start with platforms that are already available to us, such as edge-AI computers, drones, and simulated autonomous systems. The first goal is not to design a universal standard immediately. It is to find out what actually works, what fails, and what an independent human-override mechanism should look like in practice.
If the prototype proves useful, I would like to turn the results into an open test procedure that other researchers and developers can use on their own Physical AI systems.
I think this is something worth working on before autonomous AI becomes common in transportation, robotics, homes, industry, and critical infrastructure.
The main question I want to answer is straightforward:
If an AI-controlled physical system starts behaving incorrectly, becomes unresponsive, or does not follow a software-level stop command, can a human still reliably take control?
I would start by defining a small set of basic override functions. These would include emergency stopping, communication isolation, reducing the system's authority, moving it into a safe state, and recovering it afterwards.
The next step would be to build a prototype. My current idea is to use separate control hardware and a separate management channel, so that the emergency mechanism does not depend entirely on the same software or network being used by the AI system.
Then I would test it under different failure scenarios. For example, an AI agent might continue operating after a stop request, lose its main network connection, pass a task to another agent, or behave incorrectly while controlling a simulated or physical platform.
The important part is the testing. I want to measure whether the system can actually be stopped, how quickly this happens, whether any capabilities remain after intervention, and whether the device reaches a safe condition.
I plan to release the prototype code, test setup, results, and the first version of the Human Override test procedure publicly.
If the early results are good, I would like to expand the project into a broader safety-testing framework for Physical AI.
Most of the funding would go toward giving this project enough dedicated time and people to become a real research effort rather than something I work on occasionally alongside my normal academic responsibilities.
I would like to involve at least one research engineer or postdoctoral researcher, together with student research assistants. They would help with software development, hardware integration, repeated testing, data collection, and documentation.
Some funding would also be needed for edge-AI hardware, control boards, networking equipment, microcontrollers, safety controllers, and test platforms. We may also need cloud or GPU resources and access to frontier-model APIs for some of the agent-based experiments.
I would also like to use part of the funding to work with people who have stronger experience in AI safety, robotics safety, or hardware safety. I do not want the project to stay isolated inside our own lab if outside feedback can improve it.
For a larger version of the project, my rough budget would be:
PI protected research time or teaching relief: $25,000-30,000
Research engineer or postdoctoral researcher: $45,000-55,000
Student research assistants: $20,000-30,000
Hardware and Physical-AI test platforms: $20,000-30,000
Model APIs, GPU/cloud resources, and software infrastructure: $15,000-25,000
Research visits, collaboration, and technical advice: $10,000-15,000
Open-source testing, dissemination, and contingency: $10,000-15,000
I would also be happy to begin with a smaller first round if that allows us to build the first working prototype and public demonstration.
I am Muhammad Toaha Raza Khan, an Assistant Professor of Computer Engineering at Middle East Technical University.
My background is not traditional AI alignment. Most of my work has been on the systems that AI is increasingly being connected to: edge computing, vehicular networks, UAVs, IoT, non-terrestrial networks, distributed resource allocation, security, privacy, and resilient communications.
My academic profile is here:
METU AVESIS profile
My publications are here:
Google Scholar
I have published work on edge-intelligent systems, AI-based resource allocation, UAV control and networking, vehicular communications, privacy and security, non-terrestrial networks, and distributed intelligent systems. Some of this work has appeared in IEEE Transactions on Intelligent Transportation Systems, IEEE Transactions on Vehicular Technology, IEEE Sensors Journal, IEEE Communications Magazine, Information Fusion, Internet of Things, and IEEE Wireless Communications Letters.
I have also worked on funded projects related to AI resource allocation in V2X systems and post-earthquake communication resilience using UAVs and LoRa.
At METU, I supervise undergraduate research and engineering projects involving AI, computer vision, sensing, communication systems, and edge computing.
For this project, I would start with a small team built around my current research environment. I would also try to bring in collaborators with stronger direct experience in AI safety, robotics safety, or hardware safety where needed.
The most likely way this project fails is not that we cannot build anything. I think building a prototype is feasible.
The bigger risk is that we build something that works well in our own test setup but is too specific to be useful elsewhere.
Different robots, drones, vehicles, and edge systems can have very different hardware and control interfaces. It may turn out that there is no single override design that works across all of them.
Another risk is that the project stays too academic. We might publish results but fail to get other researchers, companies, or standards groups interested enough to use the test procedure.
It is also possible that software-level safety systems improve faster than expected, or that companies prefer their own proprietary emergency-control systems instead of an open approach.
If that happens, I would still want the project to produce something useful. At minimum, we should come away with evidence about which override approaches work, where they fail, and which assumptions are unsafe. The code, test cases, and results should also be reusable by other groups.
For me, producing papers alone would not count as success. I would want at least one outside group to reproduce, test, improve, or adopt some part of what we build.
For this Human Override project, I have raised $0 so far.
During the last 12 months, I secured two AdımODTÜ undergraduate research projects as project supervisor. Their total value is TRY 120,000, with one project funded at TRY 70,000 and the other at TRY 50,000. These are small university-supported undergraduate research projects and are unrelated to the Human Override proposal.
Before that, I worked on several larger funded research projects, including an International Science Partnerships Fund project on post-earthquake communication resilience, a project on AI resource allocation in mobile V2X communications, and research supported through Korean NRF and BK21 programs.
Those earlier grants are outside the last 12-month period and are not being used to fund this project