Learning From Demonstrations: Data Efficient Inverse Reinforcement Learning for Safety-Critical Dynamical Systems

About this talk

Modern infrastructure systems - including robotics and transportation - increasingly rely on automated decision making, yet they operate in regimes where exploration is unsafe and reward functions are not explicitly available. Classical reinforcement learning therefore becomes difficult to deploy: reward engineering is unreliable, and online trial-and-error is infeasible.

This talk presents a class of data-efficient inverse reinforcement learning (IRL) methods designed for safety-critical dynamical systems where only limited demonstrations and passive operational data are available. The key idea is to treat expert behavior as implicit encoding of operational objectives and constraints, and to recover structured decision policies without requiring handcrafted rewards or extensive interaction. We develop a representation-aware IRL framework that jointly learns trajectory features and reward structure while preserving computational scalability. A trajectory featurization network automatically aligns state–action sequences, removing manual feature engineering. We further derive a streamlined gradient formulation that avoids repeated trajectory back-propagation and large memory storage typical in existing IRL approaches. By computing feature expectations using one-shot expert statistics and incremental learner updates, the resulting algorithm achieves significantly improved data efficiency and numerical stability. We demonstrate the approach on benchmark tasks and autonomous driving predictions and discuss its implications for operation of autonomous vehicle systems where objectives are multi-criteria, partially observed, and difficult to specify explicitly. The broader goal is to move learning-based control from reward design toward behavior inference - enabling reliable decision-making in engineering systems where safety constraints limit exploration.