基於人類技能移轉之人工智慧晶片發展環境及其於擬人雙手機器人之應用

Intelligent robotics, Deep imitation learning, Neuromorphic computing, AI accelerators

本計畫開發能學習人類技能、具有擬人手的機器人。所發展的擬人機械手，可如人手般靈巧而快速地調適不同工作，搭配利用RRAM技術所開發的AI辨識晶片以辨識場景與物件，再透過一套虛實整合AI訓練與模擬平台，讓機器人學習人類由視覺認知到手部控制的各種技能動作。最後，經由所發展之AI晶片優化技術，將所學習到的AI模型優化後置入AI晶片中執行。整個開發流程串連人類示範動作，連結到虛擬環境訓練，產出機器人操控模型，經優化後置入AI加速晶片上執行，可大幅降低擬人機器人應用開發的門檻。

  　

This project aims to develop robots with anthropomorphic hands that can learn human skills. The anthropomorphic manipulator can adapt to different tasks as dexterously and quickly as a human hand. Its vision is enhanced with our RRAM-based AI chip to recognize scenes and objects. The robot then manipuates the objects with skills learned from human demonstration. The skills are learned through a virtual and real integrated AI training and simulation platform, which generates AI models to control the robot with visual inputs. The generated AI models are optimized through our AI chip optimization technology and finally are loaded unto the AI accelerators for execution. The entire development process starts from human demonstrations, through the virtual training platform and model optimization, and finally ends with optimized models loaded onto the AI accelerator for execution. The whole process is automated, which can greatly lower the efforts of application development for anthropomorphic robots.