Graceful Degradation in Autonomous Agents: SAGE Memory-Augmented Drone Navigation Without Language Model Dependency

Read full text Summarise with VIRAT
Authors: Ivelin Likov
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Autonomous drones that rely on language models for decision-making fail completely when connectivity is lost. This paper presents a proof-of-concept study demonstrating graceful degradation: a drone agent that continues operating at full decision rate when its LLM is disabled, using SAGE (Spatial Associative Geometric Embeddings) geometric memory as a fallback. Across all test modes (72 total agent steps), zero default decisions were recorded — the agent never stopped. SAGE-only geometric retrieval averages 2.09 seconds versus 7.49 seconds for LLM inference — a 3.6x speedup. An architectural ablation shows SAGEDivided working memory improves SAGE-only meaningful recall by 60% (5/12 to 8/12). SAGESequenceCube, a dedicated transition memory component, achieves 100% retrieval and rollout accuracy, thereby resolving the sequence-awareness gap. All experiments use text-command simulation — sensor integration is identified as future work.

Fields of Study & Subjects

AI-based Problem Solving and Planning Reinforcement Learning in Robotics Psychiatry Mental Health Neuroscience

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
23 March 2026
Language
EN
Document type
Preprint