Event
ECE Colloquium Series - Mani Srivastava, UCLA
Friday, April 10, 2026
3:30 p.m.-4:30 p.m.
Jeong H. Kim Engineering Building, Room 1110
Darcy Long
301 405 3114
dlong123@umd.edu
Speaker: Mani Srivastava, Professor, UCLA
Title: "From Models to Systems: Architecting Adaptive Compound Physical AI Across the Edge–Cloud Continuum"
Abstract: As AI moves into the physical world, the central challenge is no longer training better models, but operating complex systems of models under real-world constraints. In cyber-physical systems (CPS), application performance emerges from interactions across perception, cognition, communication, and action pipelines deployed over heterogeneous device, edge, and cloud platforms. This talk argues that performance is fundamentally a system property, not a model-level attribute.
We present a unifying framework of Compound Physical AI (CP-AI) systems, extending the traditional perception–cognition–action loop to distributed, multi-agent settings with communication and collaboration. Within this framework, system behavior is governed by a high-dimensional design space spanning model architectures, pipeline composition, deployment placement, and network dynamics. Optimizing individual components in isolation is insufficient: accuracy–latency tradeoffs, inference placement, and network variability jointly determine real-world outcomes.
Building on recent work in distributed inference, multimodal sensing, and adaptive systems, the talk highlights three key principles. First, adaptation must be system-wide, treating model selection, placement, and resource allocation as coupled control decisions across the edge–cloud continuum. Second, foundation models introduce a new layer of shared infrastructure, enabling multi-task, multi-tenant pipelines while also creating new challenges in scheduling, alignment, and control. Third, robust operation requires adaptation to both input uncertainty and resource variability, including modality quality, compute availability, and network conditions.
Through case studies in autonomous mobility and multimodal CPS, we demonstrate how these principles reshape system design and challenge conventional assumptions about model placement and optimization. The talk concludes with a research agenda toward adaptive, foundation-model-driven CPS-AI systems that continuously reconfigure themselves to meet application-level objectives such as safety, latency, and efficiency.
Bio: Mani Srivastava is a Distinguished Professor, Mukund Padmanabhan Term Chair, and Vice Chair of Computer Engineering in the Department of Electrical and Computer Engineering at UCLA, with a joint appointment in Computer Science. His research focuses on learning-enabled, resource-efficient, and trustworthy human–cyber–physical and IoT systems, spanning applications in mobile health, smart environments, national security, and sustainability. He takes a full-stack approach, addressing challenges across the edge–cloud continuum, including application design, architectural abstractions, algorithms, and platform technologies. He is a Fellow of the ACM and IEEE.
