The recent advancements of Artificial Intelligence (AI) have generated a lot of interest in the robotics community. Indeed, AI can find application in a wide variety of problems. Among these, social navigation of mobile robots is a big challenge, where ensuring non-harmful behaviors of the robotic system is fundamental. In this paper, we consider a simulated navigation problem that involves a fleet of mobile agents moving in a cross scenario, governed by a human-like behavior. With the purpose of avoiding collisions among them, we show how safe and explainable AI (XAI) methods can constitute useful tools to tailor the parameters of the behavior towards a safe, collision-free, navigation. We first explore how global native rule-based classification provides interpretable characterizations of the agents’ behavior. Afterwards, we derive safety regions, S_epsilon , denoting the zones in the parameters space where collisions are avoided, with a maximum error given by . The design of the regions is based on scalable classifiers, a technique to tune the decision function of a machine learning (ML) classifier so to bound its error on a desired class to a predefined level, combined with either probabilistic scaling (probabilistic safety regions, PSR), or with conformal prediction theory (conformal safety regions, CSR). Finally, we investigate how explainability can be provided to these regions by extracting local rules from their boundaries.
Ensuring Safe Social Navigation via Explainable Probabilistic and Conformal Safety Regions
Sara Narteni
Primo
;Alberto CarlevaroSecondo
;Maurizio MongelliUltimo
2024
Abstract
The recent advancements of Artificial Intelligence (AI) have generated a lot of interest in the robotics community. Indeed, AI can find application in a wide variety of problems. Among these, social navigation of mobile robots is a big challenge, where ensuring non-harmful behaviors of the robotic system is fundamental. In this paper, we consider a simulated navigation problem that involves a fleet of mobile agents moving in a cross scenario, governed by a human-like behavior. With the purpose of avoiding collisions among them, we show how safe and explainable AI (XAI) methods can constitute useful tools to tailor the parameters of the behavior towards a safe, collision-free, navigation. We first explore how global native rule-based classification provides interpretable characterizations of the agents’ behavior. Afterwards, we derive safety regions, S_epsilon , denoting the zones in the parameters space where collisions are avoided, with a maximum error given by . The design of the regions is based on scalable classifiers, a technique to tune the decision function of a machine learning (ML) classifier so to bound its error on a desired class to a predefined level, combined with either probabilistic scaling (probabilistic safety regions, PSR), or with conformal prediction theory (conformal safety regions, CSR). Finally, we investigate how explainability can be provided to these regions by extracting local rules from their boundaries.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.