1、具身多模态大模型-具身大脑RoboBrain.pdf

1、具身多模态大模型具身多模态大模型-具身大脑具身大脑RoboBrainRoboBrain智源 具身智能大模型研究中心王鹏伟人工智能趋势分析人工智能趋势分析1950196019701980199020002023202420252006AI 1.0AI 1.0（Small ModelsSmall Models）PerceptronPerceptronExpert SystemExpert SystemBPBP Algorithm AlgorithmXCON&MYCIN Expert Systems Become Commercially AvailableExpert SystemPC（Cost/

2、Application Scope）ChatGPTGPT-3TransformerCommercial FactorsTechinical FactorsAI 2.0AI 2.0（Large ModelsLarge Models）AI 1.0 Commercialization bottlenecksAI诞生First WaveSecond WaveThird WaveGPT-4GPT-4VGeminiSoraGPT-4oRT-2Dartmouth ConferenceEmbodied Embodied AIAI as the Next as the Next GPT Moment in AI

3、GPT Moment in AISORASORA（20242024）C ChatGPThatGPT（20232023）A AlexNetlexNet（20122012）DNNDNN（20062006）Embodied Embodied AIAI will become the core driving force of AI technology development over the next decade,leading to new revolutionary leading to new revolutionary products and reshaping industry la

4、ndscapes.products and reshaping industry landscapes.Figure from BAAI具身智能趋势分析具身智能趋势分析Single TaskSingle Ontology,Single ScenarioAI 1.0 ME 1.0Hand-Eye Coordination Robotic ArmLarge ModelBig DataRobot BrainAI 3.0+ME 3.0+General Intelligence SystemsMultiple Ontologies,Multiple ScenariosEnd-to-End Multimo

5、dal Large Model RoboticsIntelligent Robots Integrating Perception,Grasping,and Mobility2023 and Earlier2024Post-2025Multiple TasksSingle Ontology,Single ScenarioAI 2.0 ME 2.0Intelligent RobotsPerception&UnderstandingTask Decision-MakingExecution&Collaboration Evaluation&FeedbackScaling LawProven Acc

6、uracy in Large Language Models and Multimodal Large ModelsEnd-to-End Large Models in the Field of Embodied AIBasic CapabilitiesAI 1.0ME 0.0ManipulationNavigationPerception具身智能里程碑事件具身智能里程碑事件20222025SayCan2022.4GoogleRT-12022.12GooglePaLM-E2023.3GoogleVoxPoser2023.7StanfordRT-22023.7GoogleREFLECT2023.