点击论坛跳转到相应位置
基于学习的机器人操纵控制
01
论坛简介、目的与意义
02
论坛日程
03
论坛嘉宾
杨耀东 报告嘉宾、主持人
北京大学人工智能研究院研究员、博士生导师
嘉宾简介：杨耀东博士，北京大学人工智能研究院助理教授、博导，伦敦国王大学客座助理教授。科研领域包括强化学习、博弈论和多智能体系统，重点关注基于强化学习方法的群体智能涌现。他本科毕业于中国科技大学，并于帝国理工大学、伦敦大学学院获得硕士及博士学位。目前，他发表AI顶会论文及专利专著 60 余篇，谷歌引用1800余次，他的工作曾获 CoRL’20最佳系统论文奖、AAMAS’21最具前瞻性论文奖、AAAI/ACM SIGAI 优博奖参选人 (UCL唯一)、华为英国公司最佳技术突破奖、世界人工智能大会（WAIC 2022）云帆奖璀璨明星、2022年ACM SIGAI China 新星奖（Rising Star Award）。同时，他的研究受CCF-犀牛鸟基金、CAAI青年托举计划资助、人社部高层次留学人才回国资助支持。
报告题目：基于合作博弈的双灵巧手操纵控制
报告摘要：基于合作博弈的多智能体学习技术近年来取得了重要进展，尤其是在星际争霸等游戏环境中不断刷新了基线效果。但是面向智能机器人应用，例如双灵巧手的操控，合作博弈算法面临众多挑战。在此次报告中，讲者将回顾近年来合作博弈算法的发展，阐述他们的缺陷，并介绍基于置信域策略优化的多智能体强化学习算法及多智能体镜学习(Multi-agent mirror learning), 以及它们在双灵巧手控制原子动作技能学习与发育（例如，提、拔、推、拧、开）与中的进展。同时，讲者也将分享在灵巧双手操控任务中开源的Bi-DexHands框架以及基于此框架开发的安全强化学习TrustDeHands, 多任务学习、元学习任务的最新进展及展示。
李强 报告嘉宾
德国比勒菲尔德大学认知交互技术研究中心首席研究员
嘉宾简介：Dr. Qiang Li a Principle Investigator at the Center for Cognitive Interaction Technology (CITEC) in Bielefeld University, with research interests covering multimodality interaction and learning, robotic dexterous manipulation and Collaborated Robots R&D. He is also the President of Association of Chinese Computer Scientists in Germany. From 2019 to 2020, Dr. Li was a Principal Scientific Researcher and Project Leader at RoboticsX Lab. From 2012 to 2018, he was a Senior scientific researcher at Neuroinformatik group, Bielefeld University. From 2009 to 2012, he was a postdoc at CorLab, Bielefeld University. Dr. Li hold a PhD from Shenyang Institute of Automation, Chinese Academy of Sciences (2010). Dr. Li is leading the Association of Chinese Computer Scientists in Germany and developed its whole scientific organization structure and construct the Experts Committees of intelligent robots, big data and AI, Industrial IoT, autonomous driving and Brain Neurorecognition. In 2016, he founded the Sino-Germany Symposium on Intelligent Robots and developed it as a serial robotics conference (2016,2019,2020,2021) for the researchers and engineers in the whole Europe. In 2018, he was awarded “ 10 leading Chinese Talents on Science and Technology in Europe” by Federation of Chinese Professional Associations in Europe.
报告题目：Robotic dexterous manipulation: discussing from sensory-based control perspective
报告摘要：In order to autonomously work in an unstructured environment, robots have to employ their perception capabilities and robust controllers to realize the safe interaction with external environment. In this talk, I will focus on the vision and tactile sensing and integrate them into a general multi-modality closed-loop controller to implement lots of common contact-based learning and manipulation tasks. The whole talk is composed by two sub-topics:
Sub-topic 1: a general tactile servoing controller and its applications.
•exploring unknown objects
•learning robot’s body schema
•learning to use a grasped tactile tool
Sub-topic 2: a general visuo-tactile servoing controller and its applications
•unknown object’s in-hand manipulation
•learning a grasped hinged tool
董豪 报告嘉宾、主持人
北京大学前沿计算研究中心助理教授、博士生导师
嘉宾简介：董豪，现任北京大学前沿计算研究中心助理教授，博士生导师。研究围绕智能机器人中可泛化操控和自主决策展开，旨在通过生成模型和自监督学习提升。董博士在NeurIPS、ICLR、ICCV、ECCV、IROS等顶级国际会议和期刊中发表论文多篇，谷歌引用2700余次，获得ACM MM最佳开源软件奖。董博士主持或骨干参与多项国家级和省级项目。董豪本科毕业于英国中央兰开夏大学，硕士和博士毕业于英国帝国理工学院。
报告题目：可泛化的机器人操纵控制表达
报告摘要：Perceiving and manipulating 3D articulated objects (e.g., cabinets, doors) in human environments is an important yet challenging task for future home-assistant robots. The space of 3D articulated objects is exceptionally rich in their myriad semantic categories, diverse shape geometry, and complicated part functionality. Further, in practice, much hidden but important dynamic and kinematic information cannot be fully observed from visual input only, but can be queried from interactions with target objects (e.g., the axis location of a closed handleless door). Previous works mostly take only pure visual inputs, abstract kinematic structure with estimated joint parameters and part poses as the visual representations for manipulating 3D articulated objects. We design an interaction-for-perception representation to learn geometryaware, interaction-aware, and task-aware visual action affordance and trajectory proposals. In the meanwhile, to learn the non-visual but important hidden kinematic and dynamic factors, we utilise the test-time interactions to acquire non-visual information, as well as adapts the affordance prediction via proposed interactions. Our approaches have demonstrated the effectiveness in large-scale simulation dataset, and show promising generalization capabilities to novel test shapes, unseen object categories, and real-world data.
杨超 报告嘉宾
上海人工智能实验室智能决策中心青年研究员
嘉宾简介：杨超，现任上海人工智能实验室智能决策中心助理研究员、博士后，科研围绕机器人学习及应用，主要研究机器人在跨任务学习中技能的模仿、迁移与增强，在NeurIPS、AAAI、ICRA、CoRL等国际顶级学术会议与期刊上，发表多篇机器人学习与操作相关论文工作，论文在谷歌学术引用量超2000次，并带领团队于2016、2019年IROS国际机器人抓取与操作比赛中斩获多项冠军。杨超本科毕业于四川大学，硕士和博士毕业于清华大学。
报告题目：高效示教下的机器人观测模仿学习
报告摘要：回顾近十年，深度强化学习的发展当中，从Atari game到围棋对弈的AlphaGo以及最近比较火的mobile game上，深度强化学习都已经很好的解决了这一单一静态的训练化决策或博弈问题。当我们再进一步去考虑机器人应用场景的时候，其环境的复杂性动态性以及任务的多样性，均给现在的机器人学习方法提出了更高的一些要求。机器人学习问题包含示教与模仿两个阶段，但是一直存在示教样本复杂度与模仿学习效率存在平衡的挑战，针对该难题，本报告就机器人学习中的技能模仿问题，探讨高效示教下的技能观测模仿学习方法。报告将从理论角度提出了一个完整且可证明的解决方案，为高效示教的观测模仿学习提供了一个很好的方案，并实验验证该方案的可行性。
张翰博 报告嘉宾
字节跳动人工智能实验室机器人算法研究员
嘉宾简介：张翰博，现任字节跳动人工智能实验室机器人算法研究员，研究兴趣包括机器人抓取技能学习和强化学习等，主要研究内容为基于语义理解和关系理解和机器人抓取算法及其在真实复杂场景中的部署与应用，曾在RSS、IJCAI、ICRA、IROS等国际顶级学术会议与期刊上发表过多篇与机器人抓取技能学习相关的论文工作。在加入字节跳动之前，张翰博本科与博士均毕业于西安交通大学，并在新加坡国立大学完成了博士期间的联合培养学习。
报告题目：基于语言交互的场景理解与机器人抓取
报告摘要：We present INVIGORATE, a robot system that interacts with humans through natural language and grasps a specified object in clutter. INVIGORATE embodies several challenges: (i) infer the target object among other occluding objects, from input language expressions and RGB images, (ii) infer object blocking relationships (OBRs) from the images, and (iii) synthesize a multi-step plan to ask questions that disambiguate the target object and to grasp it successfully. We train separate neural networks for object detection, for visual grounding, for question generation, and for OBR detection and grasping. However, errors in visual perception and ambiguity in human languages are inevitable and negatively impact the robot’s performance. To overcome these uncertainties and simultaneously take into account the object-centric task structure, we build an object-centric partially observable Markov decision process (POMDP) that integrates the learned neural network modules. Through approximate POMDP planning, the robot separately tracks the history of observations from individual objects to update their exclusive beliefs. Besides visual observations, the robot is allowed to ask disambiguation questions when encountering semantic ambiguity and receive linguistic observations from humans. The disambiguation questions are generated based on empathy, i.e., the understanding and prediction of how humans will interpret them according to the concurrent situation. Experiments with INVIGORATE on a Fetch robot show significant benefits of this integrated approach to object grasping in clutter with natural language interactions.
