Safe Exploration of Reinforcement Learning Agents in Uncertain and Black-Box Environments

Reinforcement learning (RL) is a powerful framework for sequential decision-making. However, training and deploying RL agents in safety-critical, real-world environments remains a major challenge. Many real-world applications lack a simulator or offline dataset that captures dynamics, sensing artefacts, and long-tail edge cases well enough, so effective policies often require on-system learning, adaptation, or continual calibration after deployment. In these safety-critical environments, failures are often rare but catastrophic, exploration carries real consequences, and agents must operate under uncertainty and poorly specified safety constraints. This thesis argues that safety in RL should be understood as a system-level property that emerges from how uncertainty, risk, and constraints are handled across the entire perception–decision–control pipeline.

At the perception stage, safety-relevant outputs derived from high-dimensional sensors are treated as inherently uncertain; rather than relying on point predictions, the proposed approach attaches statistically valid uncertainty to perceptual decisions using distribution-free conformal prediction combined with lightweight stochastic sampling. This helps downstream policies distinguish well-supported regions from unreliable or ambiguous sensing.

At the decision-making stage, the thesis addresses the tendency of safety-aware RL methods to become either over-confident or uniformly pessimistic. An uncertainty-aware safety evaluation mechanism is introduced to separate expected safety outcomes from uncertainty about those outcomes, enabling selective conservatism: the agent can behave cautiously when predictions are poorly supported while avoiding unnecessary avoidance in regions where experience provides reliable safety information.

Finally, at the control stage, the thesis considers environments in which system dynamics and safety constraints are not available, motivating a post-hoc enforcement mechanism that does not assume access to an analytic model or an explicit cost function. A data-driven shielding layer is developed that learns from experience to distinguish safe from unsafe behaviours in continuous black-box settings, and uses this learned notion to filter, correct, or veto candidate actions before execution.