Robots are increasingly capable because they learn from data, but a learned policy comes with no promises: it can behave unpredictably just outside the situations it was trained on, and there is rarely a way to state what it will never do. Formal methods offer the reverse trade-off. They can establish provable correctness, but they need a model, and specifications are usually written over a fixed, finite set of propositions. My research sits at the interface: first-order temporal specifications and situation-calculus programs that quantify over objects rather than propositions, and the question of which guarantees survive when parts of a system are learned. I work on both the theory and the deployment on real robots.
Research threads
Specifications that generalize
First-order temporal specifications and procedural (Golog / situation-calculus) programs as the interface between human intent and machine behavior. Written once, they hold for any number of objects, products, and instance sizes.
Guarantees under learning
Formal methods presuppose a model; learned components do not come with one. This thread asks which correctness guarantees can still be established when parts of an autonomous system are learned rather than specified.
Robots that actually run
Long-horizon task and motion planning, execution monitoring, goal reasoning, and deployment on physical robots in industrial settings. This is where the theoretical questions come from in the first place.
News
- 6 July 2026
“WorkBenchMark: A LEGO-based Assembly Benchmark with an Assembly-by-Disassembly Baseline for the Smart Manufacturing League” (with Wenbo Ma, Daniel Swoboda, and Matteo Tschesche) was presented at the RoboCup 2026 Symposium in Incheon, South Korea (project page).
- 28 June 2026
“Learning Bilevel Policies over Symbolic World Models for Long-Horizon Planning” (with Dillon Z. Chen, Toryn Q. Klassen, and Sheila A. McIlraith) and “Learning General Policies for Partially Observable Deterministic Planning” (with Samridhi Kalra and Hector Geffner) were both presented at the ICAPS Workshop on Generalized Planning (GenPlan) in Dublin.
- 23 January 2026
Our paper “Satisficing and Optimal Generalised Planning via Goal Regression” (with Dillon Z. Chen, Toryn Q. Klassen, and Sheila A. McIlraith) was presented at AAAI 2026 in Singapore (proceedings, presentation).