research

overview of my research and selected projects

An overview of my research for developing safe and explainable robots that interact with humans.

NEURO-SYMBOLIC AI

My work in neuro-symbolic AI combines temporal logic with learning-based control so robots can solve long-horizon tasks more safely, reliably, and interpretably.

Left: Reward modeling and prediction via Gaussian processes and deep neural networks. Right: Extracted specification-consistent behaviors in simulations, including Nvidia Isaac.

Modeling and control synthesis for long-horizon tasks

I combine Behavior Trees with temporal logics and timed automata to support formal verification, inconsistency detection, and control synthesis for complex robotics tasks (Matheu et al., 2025); (Matheu et al., 2025); (Puranic et al., 2026).

Learning reward functions and policies from demonstrations

I develop neurosymbolic learning-from-demonstrations methods that infer temporal reward functions and policies from demonstrations, improving safety and performance in settings where standard inverse RL is insufficient (Puranic et al., 2021); (Puranic et al., 2021); (Puranic et al., 2023); (Puranic et al., 2026).

Improving performance beyond the demonstrator

I study how temporal-logic-guided reinforcement learning can help robots self-monitor and improve beyond the quality of the demonstrations they receive, including through risk-sensitive methods that reduce overestimation bias and improve safety (Puranic et al., 2023); (Puranic et al., 2024); (Enwerem et al., 2025).

INTERPRETABLE/EXPLAINABLE AI (xAI)

My work on explainable AI focuses on turning demonstrations and time-series data into human-understandable models of behavior and system constraints.

Left: Generating graphs that explain demonstrator performance and formal specification conflicts. Center: Neural reward modeling from inferred graphs. Right: Mining formal specifications from time-series data.

Generating explainable temporal-logic graphs

I introduced Performance Graph Learning (PeGLearn) to infer explainable temporal-logic structures from human demonstrations, reducing manual effort while improving interpretability and supporting human-centered evaluation (Puranic et al., 2023).

Mining specifications from temporal data

I develop methods to infer formal specifications from time-series data and to explain anomalies in cyber-physical systems, making complex AI-driven behavior easier to diagnose and verify (Mohammadinejad et al., 2020); (Mohammadinejad et al., 2020); (Noorani et al., 2026).

COMPUTER VISION

My computer vision work focuses on formal evaluation of perception systems and on developing vision-based metrics for assessing real-world performance.

Left: Timed Quality Temporal Logic (TQTL) for specifying spatio-temporal properties of perception systems. Right: Vision-based estimation of needle-entry deviation during robotic suturing.

Evaluating perception algorithms

I introduced Timed Quality Temporal Logic (TQTL) to specify spatio-temporal properties of vision-based perception systems without requiring extensive ground-truth labels (Balakrishnan et al., 2019).

Measuring surgical performance from vision

I developed a vision-based approach for estimating needle-entry deviation during robotic suturing, offering a proxy for tissue handling quality when force feedback is unavailable (Puranic et al., 2019).