Orion framework enables self-adaptive memory for on-device continual learning

Online continual learning (OCL) allows AI models to adapt in real time, making it crucial for dynamic robotic applications. However, deploying OCL on resource-limited devices is challenging because memory constraints create trade-offs between training latency, plasticity, and stability. Traditional offline parameter tuning fails to account for the shifting memory pressure and workload complexity as OCL progresses. To solve this, researchers from UC Irvine (Zexin Li, Nikil Dutt, Cong Liu) introduce Orion, a holistic framework that co-optimizes these three factors under strict memory limits.

Orion's core innovation is URGE, a unified runtime indicator grounded in the "Buckets effect"—the idea that system performance is bounded by its scarcest resource. URGE dynamically reallocates memory across OCL components by jointly coordinating batch processing, replay buffers, and optimization strategies at both the OS and application levels. The framework also adds system-level data prefetching to maximize efficiency. Implemented using the Avalanche-lib library, Orion was tested across multiple OCL algorithms, benchmarks, and hardware platforms common in autonomous robotics. In a realistic autonomous navigational robot, Orion delivered significant training speedups while maintaining balanced performance and adapting to various scenarios, all with minimal runtime, memory, and energy overhead.