Learning to predict and avoid threat is fundamental for survival. One key behavior is avoidance — learning to perform certain actions to prevent impending aversive outcomes, prior to their occurrence. Maladaptive avoidance is a crucial feature of anxiety and trauma-related disorders. Hence, unraveling the mechanisms of avoidance learning might have real-world implications. While the boundary conditions of classical (Pavlovian) conditioning in humans are well-known, avoidance learning relies on distinct neurobiological mechanisms and potentially follows different rules. Previous work has largely relied on tasks employing a small number of unnatural actions. Here, we set out to establish boundary conditions of avoidance learning in a naturalistic task. We sought to confirm or reject canonical learning phenomena known from classical aversive conditioning, focusing on second-order conditioning, overshadowing, and partial reinforcement. In three experiments (N = 100 each), we used a virtual reality platform (CogLearn toolkit for Unity) with three-dimensional coloured objects as conditioned stimuli (CS) and an uncomfortable sound as unconditioned stimulus (US), allowing for unconstrained and uninstructed avoidance actions. Our primary outcome was the mean distance from the CS during the CS presentation (i.e. before the US). We found strong evidence for second-order-conditioning (larger avoidance responses for second-order CS+ compared to second-order CS-), overshadowing (cue-competition in a compound stimuli), partial-reinforcement-acquisition-effect (slower avoidance learning for partially CS+ compared to fully reinforced CS+), and partial-reinforcement-extinction-effect (slower extinction for partially reinforced CS+ compared to fully reinforced CS+). Our findings identified four canonical learning phenomena that are diagnostic for building cognitive-computational models of avoidance learning.