Research Analyzer
← Back ICRA 2026

A Neuro-Inspired Control Architecture to Enhance Robot Self-Preservation and Adaptation in Autonomous Navigation Tasks

Andrea Usai, Alessandro Rizzo

PDF

AI summary

Key figure (auto-extracted from paper)
A neuro-inspired architecture dynamically tunes controller weights via a reinforcement-learning amygdala, significantly improving robot safety, adaptability, and computational efficiency in unpredictable environments.
Neuro-inspired control Robot self-preservation Soft Actor-Critic Nonlinear MPC Adaptive navigation Bioinspired robotics

Problem

Standard model-based and learning-based navigation methods struggle with manual weight tuning, poor generalization, and inadequate self-preservation in complex, dynamic environments.

Approach

Inspired by the neuroscience Low Road fear pathway, the system uses a reinforcement learning agent to assess environmental danger and dynamically adjust the weights of a nonlinear model predictive controller for adaptive navigation.

Key results

  • Superior obstacle clearance and safety in static and dynamic scenarios
  • Enhanced adaptability and generalizability across varying risk levels
  • Improved computational efficiency compared to standard NMPC and APF
  • Novel real-time reinforcement learning mechanism for NMPC weight tuning

Why it matters

Enables safer, more autonomous robots for critical real-world applications like search-and-rescue and infrastructure inspection where self-preservation is paramount.

Abstract

Ensuring survival and self-preservation is essential to design intelligent robots that adapt to dynamic and unfamiliar environments. Inspired by the dual-pathway model from neuro- science, we introduce a control architecture designed to ensure the adaptability of robotic behavior during navigation. This ap- proach parallels the neuroscientific “Low Road” paradigm by incorporating constructs resembling the thalamus, implemented as a nonlinear filter; the amygdala, modeled as a Soft Actor-Critic (SAC)reinforcementlearningagent;andthebrainstem-cerebellum connection, represented by a Nonlinear Model Predictive Con- troller (NMPC). Our findings indicate superior adaptiveness, gen- eralizability, and computational efficiency compared to standard NMPCs and Artificial Potential Fields in both static and dynamic environments with obstacles of varying risk levels.

Index terms

Bioinspired Robot Learning Neurorobotics Cognitive Control Architectures

Related papers