Selected Projects
work
ICML 2026
Constrained Decoding for Robot Foundation Models
Enforces safety specifications expressed as Signal Temporal Logic (STL) formulas on transformer-based robot foundation models at inference time — no retraining required.
RSS 2025
ViSafe — Vision-enabled Safety for High-speed Detect and Avoid
Vision-based control barrier functions for UAV safety validated via sim-to-real transfer from Isaac Sim to field deployment, achieving 71% improvement over baselines across 70+ flight hours.
RA-L / ICRA 2026
STLCG++ — Differentiable Signal Temporal Logic for Robot Learning
A PyTorch/JAX library integrating differentiable temporal logic into robot learning pipelines, delivering 100× speed-ups and outperforming baselines for trajectory optimization and diffusion-policy case studies.
NFM 2024
Safe Planning via Incremental Decomposition of Signal Temporal Logic
Devised a requirement decomposition framework for integrated task and motion planning, reducing solve time by 65% by combining correctness of specification-guided synthesis with efficient convex optimization for trajectory synthesis.
Under Submission
ETL: Runtime Monitoring via Embedding Temporal Logic
A temporal logic framework for runtime monitoring of perception-based autonomous systems directly in learned embedding spaces, without discrete state abstraction.
Robot Foundation Models for Navigation and Manipulation
Pretrained multi-modal robot foundation models via large-scale behavior cloning for mobile manipulation and navigation, with General Robotics.
ICSE 2025
ATLAS: Learning Formal Behavior Rules from Robot Demonstrations
Built a framework for learning formal LTL behavior specifications from robot demonstrations using MaxSAT, enabling structured policy synthesis with user-defined constraints.
ICRA 2023
Online STL Tree Search for Guided Imitation Learning
Employed Monte Carlo Tree Search (MCTS) as a means of integrating STL specification into a vanilla LfD policy to improve constraint satisfaction.
FM 2024
RL Robustness Analysis
Analysing robustness of safe reinforcement learning policies and control agents in the face of environmental deviations such as steering, friction, and sensor noise.
Trust Elicitation and Restoration in Assistive Robots
Pilot user study investigating the impact of policy customization and perspective on perceived trust in RL-based assistive robot policies.
Needle
A comprehensive deep learning library from scratch, enabling GPU acceleration, automatic differentiation, and customizable layers, loss functions, and optimizers as part of DLSystems course project.