Cross-Layer Adaptive Co-Design for Mission-Critical Embodied Intelligence
Rajat Bhattacharjya
Department of Computer Science, University of California, Irvine
ACM/IEEE ESWEEK PhD Forum, Barcelona, Spain, October 2026.
Abstract
Mission-critical embodied agents must sustain reliable perception and decision-making as missions, sensing conditions, connectivity, resources, and available human support change. Yet existing systems typically optimize isolated algorithms or execution choices, making otherwise capable pipelines brittle when these factors vary jointly. My research develops cross-layer adaptive co-design methods that coordinate algorithmic, architectural, communication, and runtime decisions according to mission and operating state. Across my completed works, adaptation moves progressively closer to the physical decision—from making computation feasible, to invoking it when valuable, matching execution to mission-required evidence, and validating whether prior actions remain safe to reuse—yielding up to 93.98% lower energy and 76.6% fewer safety violations. Building on this foundation, my dissertation advances a central premise: rather than directly selecting a model or system configuration, an agent should first express the next physical decision as an evidence contract specifying the information, fidelity, freshness, uncertainty, deadline, and risk it requires. A bounded metareasoning runtime will then compose computation, communication, validation, and assistance strategies to satisfy that contract according to their expected decision value, cost, and validity, and verify satisfaction before action. The intended outcome is a unified runtime that closes the loop from mission intent to sufficient evidence to trustworthy physical action.