專題演講

摘要 (Abstract)：

    本次演講主題是一種基於生成式AI的創新機器人認知決策系統，旨在提升機器人如同人類大腦般的學習、推理和決策能力。傳統上，機器人依賴預先編程或機率模型進行決策，但本系統著重於模擬人類的認知過程，包含感知、記憶和假想生成，核心是一個基於神經元的假想生成模型 (Hypothesis Generation Model, HGM)。

    為了克服傳統學習系統中不穩定的收斂問題，本系統引入了記憶模型 (Memory Model)，包含短期記憶和長期記憶，參考認知心理學的記憶理論，以更準確地評估學習結果並產生改進的假想生成，從而實現更穩定的學習過程和提升決策效能。此外，感知模型 (Perception Model) 利用編碼器將環境資訊轉換為易於處理的感知編碼，有助於系統找到最佳解決方案。在研究中以中文書法為驗證應用，展示了虛擬和真實機器人系統在學習複雜動作和呈現書法美感方面的有效性和穩定的學習收斂性。

摘要 (Abstract)：

   The rapid evolution of Earth observation satellites is driving a transformation from passive sensing platforms to intelligent, autonomous spaceborne systems. Traditionally, satellites capture raw imagery and transmit massive volumes of data to ground stations for post-processing. However, the increasing demand for high-resolution video, real-time analytics, and time-sensitive applications—such as maritime surveillance and disaster monitoring—has exposed fundamental constraints in onboard power, computational resources, and communication bandwidth. These challenges require a shift toward energy-efficient edge intelligence directly implemented in orbit.

This talk presents an IC design-based approach to enabling AI at the edge of space. By integrating hardware–software co-design principles with FPGA- and accelerator-oriented architectures, we demonstrate how real-time video compression and AI inference can be achieved within strict power envelopes. Key design strategies include parallel processing pipelines, quantization-aware neural network optimization, memory-efficient dataflows, and low-power arithmetic architectures tailored for spaceborne environments. From circuit-level considerations to system-level integration, energy-adaptive computation mechanisms dynamically balance precision, throughput, and power consumption according to mission requirements.

Special emphasis is placed on lightweight neural network architectures for small-object detection in high-altitude imagery, where targets may occupy only a few pixels. Through hardware-aware model compression and efficient mapping onto reconfigurable logic or custom IC accelerators, satellites can extract mission-critical information onboard, significantly reducing transmission load while enabling event-driven responses.

By embedding intelligence into the sensing pipeline through IC design-based acceleration, satellites evolve from simple data collectors into cognitive systems capable of adaptive perception and onboard decision-making. AI at the edge of space thus represents a critical advancement toward autonomous, efficient, and responsive next-generation satellite platforms.