Cognition Envelopes for Bounded AI Reasoning in Autonomous UAS Operations

TL;DR

The paper addresses the risk that LLM/VLM-based reasoning in autonomous CPS, especially sUAS SAR, can produce flawed or unsafe decisions due to hallucinations and misalignment. It introduces Cognition Envelopes as external guardrails that constrain reasoning outcomes; the concrete instantiation uses a pSAR probabilistic envelope and a Mission Cost Evaluator to monitor CAP outputs. Key methods include modeling the Probability of Area with $p(c) = R(c) \cdot A(c)$, where $R(c)$ is reachability and $A(c)$ is environmental affinity built from 11 features using radial-basis functions, plus evidence updates and entropy-based decision thresholds $H_{\text{norm}}$ to balance exploitation and exploration. Experiments with vignette-based SAR scenarios yield about 95% CAP relevance accuracy and show that updating the POA after clue discovery increases plan approvals, underscoring the practical value and outlining open software-engineering challenges for broader deployment.

Abstract

Cyber-physical systems increasingly rely on Foundational Models such as Large Language Models (LLMs) and Vision-Language Models (VLMs) to increase autonomy through enhanced perception, inference, and planning. However, these models also introduce new types of errors, such as hallucinations, overgeneralizations, and context misalignments, resulting in incorrect and flawed decisions. To address this, we introduce the concept of Cognition Envelopes, designed to establish reasoning boundaries that constrain AI-generated decisions while complementing the use of meta-cognition and traditional safety envelopes. As with safety envelopes, Cognition Envelopes require practical guidelines and systematic processes for their definition, validation, and assurance.

Figures (7)

• Figure 1: Layered decision-making guardrails: the outer Safety Envelope sets hard limits, while the Cognition Envelope applies external guardrails on inner-level decisions.
• Figure 2: An sUAS discovers a backpack at the trailhead while conducting a trail-based search for a lost hiker. This clue could trigger mission-level adaptations.
• Figure 3: The Clue Analysis Pipeline includes four stages. The VLM generates a caption for the clue, while RAG + LLM analyzes its relevance. The Task Planner and Triager then decide whether to enact, revise, or defer the resulting action to a human operator. The Cognition Envelope approves the task or redirects it to a human for further assessment.
• Figure 4: Decomposition of Cognition Envelope requirements showing requirements related to probabilistic coherence (pSAR) and Mission Cost Evaluator (MCE)
• Figure 5: A Vignette establishes the context for each SAR contextualized decision point within our experiments. The vignette includes a lost person profile, environment, and clue data such as an image frame and the clue location. This example shows baseline information used for Test 2 and additional clues used in its variants.