UC Riverside

Safety-critical cyber-physical systems (CPS), such as autonomous vehicles and robots, increasingly rely on deep neural networks. While these systems require high prediction accuracy, they must simultaneously satisfy strict end-to-end timing constraints under tight memory and energy budgets. The challenge extends beyond static inference: realistic deployments face non-stationary data and fluctuating power conditions, necessitating online adaptation systems. However, existing machine learning methods largely overlook the timing, memory, and accelerator-contention constraints inherent to embedded CPS, resulting in unpredictable latencies and frequent out-of-memory failures.This dissertation addresses the gap between algorithmic machine learning and real-time systems by developing a predictable and self-adaptive On-Device machine learning stack. We demonstrate that predictable learning is a distinct systems problem requiring cross-layer coordination rather than isolated algorithmic adjustments.Specifically, the dissertation develops three tightly coupled thrusts. First, to make modern models runnable On-Device under stringent memory constraints, it exploits multi-input/multi-output structure, weight sharing, and training–inference co-design so that multi-task perception, reinforcement learning, and continual learning workloads fit within small-device memory budgets while remaining schedulable as modular units (MIMONet, MixTraining, R3 , Orion). Second, to reduce latency and deadline misses in real-time learning pipelines, it shapes end-to-end deadlines across perception–control DAGs, enforces intermediate timing contracts, and designs runtimes that safely overlap evaluation, inference, and in-the-loop training on shared accelerators (RED, AOCL). Third, to enable deployable learning ecosystems in uncertain, resource-constrained environments, it introduces environmentand energy-aware controllers that adapt sampling, buffering, and update cadence online as workloads, data distributions, and power budgets evolve, sustaining timing and accuracy under real-world drift (Genie, EOCL).Ultimately, this work provides the systems foundation required to move beyond “deploy then freeze” edge AI, enabling continuous, resource-aware, and predictable adaptation for mission-critical robotics.