Evolution of Autonomous Systems: Core Architectures and Taiwan's Unique Advantages

Over eons, synaptic connectivity endowed animals’ carbon-based brains with high-dimensional processing for survival and decision-making. We now approximate the perceive–understand–decide pipeline with ANNs and multimodal models, while modern silicon handles large, high-dimensional data within explicit energy and latency budgets.

Building truly human-like autonomous systems still faces critical bottlenecks, demanding disruptive architectures to maintain consistent performance in dynamic, unpredictable environments on a few-tenths-of-a-watt power budget.

Many autonomous systems still center on a compute-intensive, centralized “System 2,” paired with a low-latency sequence-generation engine (the so-called “System 1”). While fine for labs or simulations, this design makes balanced trade-offs among reliability, size, mass, and energy efficiency nearly impossible; the absence of proprioception impedes long-horizon consistency, robustness, and adaptability.

Shift from centralized compute-first to layered, bottom-up autonomy centered on distributed proprioception: low-latency, low-power perception–action loops (the real System 1) handle immediate responses; a predictive, adaptive brain (System 2) governs strategy. Evolutionary distillation and related methods optimize each subsystem’s energy efficiency, mass, volume, reliability, and safety.

Leveraging strengths in semiconductor manufacturing, advanced packaging, and embedded-system integration, Taiwan is well positioned to drive and define the low-power, high-volume sensing and actuation chiplet markets, building up a complete supply chain of “neural reflex systems” (low-latency, real-time reaction subsystems) for the autonomous systems and robotics industry. This will not only reduce dependence on ultra-high-compute “brains” but also strengthen Taiwan’s position in the global ecosystem ahead of large-scale commercialization of autonomous robots and intelligent systems, and enhance its strategic influence in future end-to-end system competition.

經過漫長的演化，動物的大腦靠突觸把神經細胞連起來，慢慢長出「高維度資訊處理」的能力，才能在複雜環境中存活與決策。對應到今天，我們用人工神經網路（ANNs）與多模態模型，在功能層面近似「感知→理解→決策」的流程；同時，現代矽基硬體也能在明確的能耗與延遲預算內，處理大量高維度資料。

但要做出真正「類人」的自主系統，還卡著關鍵瓶頸：要在動態、難預測的場景裡長時間保持穩定表現，而且功耗只有二十多瓦，必須採用更顛覆的新架構。

目前許多自主系統仍以計算密集、集中式的「System 2」為核心，再配一個低延遲的序列生成引擎（所謂「System 1」）。這在實驗室或模擬環境或許可行，但一到實務現場，就很難同時兼顧可靠性、尺寸、重量與能效；而且欠缺本體感知（proprioception），使得長期一致性、穩健度與適應力都受限。

解法是從「算力集中優先」轉到「由下而上、以分散式本體感知為核心」的分層自主架構：由低延遲、低功耗的感知—動作融合迴路（真正的 System 1）負責即時反應；上層由更會預測、能調適的「大腦」子系統（System 2）負責策略治理。配合演化式蒸餾等方法，各子系統可分別在能效、質量、體積、可靠性與安全性上做到最佳化。

台灣的優勢在這裡特別關鍵：我們結合半導體製造、先進封裝與嵌入式整合，有機會主導低功耗、高出貨量的感測與致動 chiplet（小晶片）市場，建立面向自主系統與機器人產業的「神經反射系統」（低延遲、即時反應子系統）完整供應鏈。這不只減少對超高算力「大腦」的依賴，也能在大規模商用化之前，強化台灣在全球生態系的地位，並提升未來端到端系統競爭中的戰略影響力。