Tailored Computing: Cross-Layer System, Architecture, and Silicon Co-Design for Physical Intelligence
Author(s)
Wan, Zishen
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Physical intelligence -- where embodied agents perceive, reason, plan, and act in the physical world -- is emerging as the next transformative computing platform, offering significant potential impact across robotics, healthcare, manufacturing, and scientific automation. Yet today’s physical intelligence systems suffer from prohibitive latency, excessive energy consumption, and fragile reliability when deployed on resource- and power-constrained platforms operating in dynamic, uncertain environments. The fundamental mismatch between advanced algorithmic intelligence and the underlying computational substrate limits scalability, robustness, and real-world deployment.
This dissertation addresses this gap through cross-layer system-architecture-silicon co-design for physical intelligence, integrating system runtimes, heterogeneous compute substrates, and solid-state silicon. These works develop tailored computing architectures that unify high-level cognitive reasoning with low-level perceptual actuation, realized through algorithm-software co-design, technology- and integration-driven cognitive architectures, and programmable System-on-Chip (SoC) tapeouts. These contributions collectively deliver substantial improvements in real-time responsiveness, multi-agent scalability, and energy efficiency, validated through hardware simulations, FPGA prototypes, GPU-resident heterogeneous systems, and fabricated SoC silicon.
The research advances the field through three synergistic thrusts:
First, to enable efficient and scalable high-level cognitive reasoning, this work introduces the end-to-end system-architecture-silicon co-design frameworks for neuro-symbolic AI. It presents the CogSys and REASON architectures featuring unified intermediate representations, flexible microarchitectures, and scalable dataflow to accelerate neuro-symbolic workloads. These innovations culminate in the design and fabrication of the programmable heterogeneous SoC for neuro-symbolic cognition. The resulting silicon prototype demonstrates how memory-centric datapaths and software-defined power management can deliver real-time reasoning with orders-of-magnitude energy-efficiency improvements over commodity hardware.
Second, to advance efficient and scalable low-level perceptual autonomy, this dissertation bridges the divide between deliberative planning and reactive control. It introduces ReCA, an integrated architecture for cooperative embodied agents that unifies fast and slow thinking loops, and MulBERRY, an energy-aware framework for autonomous swarms. These systems optimize the “sense-plan-act” loop via hierarchical coordination and hardware acceleration for kernels such as SLAM and path planning, enabling robust and scalable multi-agent deployment in dynamic environments.
Third, to achieve energy- and safety-aware autonomous operation, this work develops an end-to-end cross-layer reliability modeling infrastructure that exposes inherent robustness variations within autonomous systems. Building on these insights, it proposes a vulnerability-adaptive protection paradigm that dynamically allocates protection budgets proportional to kernel robustness. This strategy achieves high functional safety coverage with minimal overhead, an essential requirement for embodied agents operating in unpredictable, resource-constrained settings.
Bridging computer architecture, systems, and silicon, this dissertation advances the computational foundations needed for physical intelligence -- enabling embodied agents that can think and act efficiently, adaptively, and reliably in the real world.
Sponsor
Date
2025-12
Extent
Resource Type
Text
Resource Subtype
Dissertation (PhD)