# Bioinspired Stimulus Selection Under Multisensory Overload in Social Robots Using Reinforcement Learning

## TL;DR

This paper introduces a bioinspired system using reinforcement learning to help social robots manage sensory overload by prioritizing relevant stimuli in real time.

## Contribution

A novel reinforcement learning-based attentional system inspired by neurocognitive mechanisms to manage stimulus prioritization in social robots.

## Key findings

- The system effectively modulates sensory signals and reduces redundant inputs in overstimulating scenarios.
- Compared to a baseline queue-based method, the system improves expression management by reducing queue size and delay.
- Three case studies demonstrate the system's ability to enhance stimulus selection in real-world robot interactions.

## Abstract

Autonomous social robots aim to reduce human supervision by performing various tasks. To achieve this, they are equipped with multiple perceptual channels to interpret and respond to environmental cues in real time. However, multimodal perception often leads to sensory overload, as robots may receive numerous simultaneous stimuli with varying durations or persistent activations across different sensory modalities. Sensor overstimulation and false positives can compromise a robot’s ability to prioritise relevant inputs, sometimes resulting in repeated or inaccurate behavioural responses that reduce the quality and coherence of the interaction. This paper presents a Bioinspired Attentional System that uses Reinforcement Learning to manage stimulus prioritisation in real time. The system draws inspiration from the following two neurocognitive mechanisms: Inhibition of Return, which progressively reduces the importance of previously attended stimuli that remain active over time, and Attentional Fatigue, which penalises stimuli of the same perception modality when they appear repeatedly or simultaneously. These mechanisms define the algorithm’s reward function to dynamically adjust the weights assigned to each stimulus, enabling the system to select the most relevant one at each moment. The system has been integrated into a social robot and tested in three representative case studies that show how it modulates sensory signals, reduces the impact of redundant inputs, and improves stimulus selection in overstimulating scenarios. Additionally, we compare the proposed method with a baseline where the robot executes expressions as soon as it receives them using a queue. The results show the system’s significant improvement in expression management, reducing the number of expressions in the queue and the delay in performing them.