报告摘要:
With the advance in optimal control and reinforcement learning, latestcontrollers have demonstrated exceptional performance on complex multi-
limbedrobots, including the application in aerial manipulation, quadrupedal locomotion,and dexterous hand skills. Despite these advancements, comprehensive safetyvalidation remains a prerequisite for their large-scale real-world deployment.Current state-of-the-art (SOTA) controllers exhibit robustness againststandard testing paradigms, domain randomization, and evaluations by humanexperts. In this talk, I will first introduce our progress on enhancingdexterity of the robots, and then our recent research revealing vulnerabilitiesin long-tested state-of-the-art controllers on the legged locomotion anddexterous manipulation when subjected to minor perturbations generatedby AI agents. This study highlights critical safety concerns and emphasizesthe necessity of addressing these vulnerabilities to enhance system reliability.The failure cases identified in our analysis offer valuable insights intosystem components, providing a foundation for improving the robustnessand safety of black-box neural controllers.
报告人简介:
Fan Shi is an Assistant Professor in National University ofSingapore. Before that, he was a Postdoc Researcher working with Prof.Marco Hutter and Prof. Stelian Coros in ETH Zurich. His research focusedon learning and control for multi-limbed robots in locomotion and manipulation.He obtained his Master and Ph.D. degree from the Univ. of Tokyo supervisedby Prof. Masayuki Inaba, and Bachelor degree from Peking University. Hegot several highly competitive awards, such as the Dean’s Award for hisDoctoral Thesis, IEEE RAS/JJC Young Award in ICRA 2020, AI Safety FinalistAward in Switzerland.