Sayan Mitra
University of Illinois Urbana-Champaign, IL·Computer Science
Self-driving cars and autonomous drones now rely on deep neural networks to interpret camera images, deciding where the lane is or where the runway lies. That creates an awkward problem for safety: control engineers can prove a steering law is safe if they know how accurate the position estimate is, but a neural network's accuracy is not something anyone has written down in a form a proof can use.
- Fields
Computer Science
Engineering
Mathematics & Statistics
- Topics
perception contracts
operating design domain
formal verification
- Grants
- 1 known grant, through Jun 2029
What they work on
- Develop refinement algorithms that synthesize perception contracts, machine-checkable bounds on how wrong an AI perception module can be, together with the operating design domains in which they hold, for pipelines that fuse camera and inertial sensor data.
- Extend contract-based verification beyond simple invariance (never leaving a safe set) to history-dependent correctness properties, and develop theory and algorithms for monitoring contracts at runtime.
- Test whether generative AI techniques can discover operating design domains that capture the variation present in real-world scenes, rather than hand-specified environmental assumptions.
- Integrate perception contracts and generated operating domains into a control design method built on an extension of robust control barrier functions, certificates that keep a system within a safe region despite disturbances.
- Evaluate the framework on agile flight and automated driving using photorealistic simulation and physical vehicles, including a vision-based autonomous racing benchmark released to the community.
More about the research
The assumption gets buried informally, and the question "under what conditions can this system be trusted?" has no rigorous answer. This project proposes a way to make those conditions explicit, computable, and useful for designing systems rather than left implicit.
Why it matters
Autonomous vehicles and aircraft are hard to certify because no one can state precisely when their AI perception can be relied upon. Making those operating conditions formal and checkable could give engineers explicit design boundaries, give certification authorities a rigorous basis for approval, and support runtime monitors that detect when a vehicle has left the conditions it was certified for. Benefits could extend to transportation, agriculture, public safety, and defense applications that depend on trustworthy autonomy. Released benchmarks would also let other researchers compare methods.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Somil Bansal
Stanford University, CA·Engineering·Early-career
Develop training techniques that make machine-learning perception policies robust by construction, so camera-based controllers stay safe in new situations.
Radoslav Ivanov
Rensselaer Polytechnic Institute, NY·Computer Science
Develop extrinsic neuro-symbolic calibration that adjusts the uncertainty of already-trained neural networks to match the assumptions of downstream state estimators and controllers, without retraining the networks.
Qi Chen
University of California-Irvine, CA·Computer Science·Early-career
Develop system-to-AI and AI-to-system mapping methods that generate physical-layer attacks on autonomous vehicle AI while keeping them physically realizable and semantically equivalent across diverse real system designs.
The school
University of Illinois Urbana-Champaign: 41 labs on PI Pages, $907M a year on research, 870 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.