Safety first: Toward safe action selection with contextual affordances

Reinforcement Learning empowers agents to make decisions in environments with the aim of maximizing a defined reward. In the area of robotic exploration, the challenge lies in enabling the agent to navigate its surroundings while avoiding dangerous states. Contextual affordance, a predictive model, anticipates the consequences of actions based on the agent’s current state. This model proves invaluable in guiding the agent away from dangerous situations. In the pursuit of enhancing robotic exploration safety, this paper examines the efficacy of a robotic agent employing contextual affordance to steer clear of unsafe states during exploration. The evaluation encompasses both a controlled environment and an uncontrolled setting. The results obtained show the learning agent is able to avoid dangerous states more effectively when using contextual affordance. By contrasting the agent’s performance in these scenarios, our study reveals a pronounced improvement in the agent’s ability to avoid unsafe states when leveraging contextual affordance.