Continual Learning of Foundation Models for Robotic Navigation

Introduction

Conventional learning methods often struggle in dynamic settings because agents are usually trained for fixed tasks and distributions. Recent studies describe continual learning (CL) of foundation models as a promising direction for enabling agents to adapt to new tasks while preserving prior capabilities. This is especially relevant for robotic navigation, where environments, goals, and sensory conditions evolve over time.

 Objectives

·  Study continual learning of foundation models for incremental navigation tasks.

·  Compare traditional methods against CL approaches.

·  Evaluate retention, transfer, and sample efficiency.

Tasks

1.     Literature Review: CL, CL of foundation models, and robotic navigation.

2.     Simulation Setup: Create a sequence of navigation tasks with increasing variation in maps, obstacles, or goals.

3.     Method Development: Implement a baseline agent and two or more state-of-the-art CL methods in navigation.

4.     Evaluation: Measure success rate, forgetting, adaptation speed, and data efficiency.

5.     Optional Extension: Add safety constraints or knowledge-guided components.

Pre-requisites

Python, continual learning, simulation experience is a plus.

Work

20% Theory, 60% Programming/Simulations, 20% Writing

Contact

Ali Sabzi Khoshraftar (a.sabzikhoshraftar@utwente.nl)