Cosmos 3: Omnimodal World Models for Physical AICosmos 3 unifies language, image, video, audio, and action into an open world-model backbone for physical AI.
NVIDIA2026World ModelsPhysical AISynthetic DataPrimary-source guide
Research archive
Landmark Physical AI papers with plain-language explanations, prerequisites, impact, and direct primary-source links.
PaperOrganizationYearConcepts
NVIDIA Isaac GR00T N1.7GR00T N1.7 is an open vision-language-action model for generalized humanoid manipulation skills.
NVIDIA2026VLAHumanoidsDeploymentGemini Robotics-ER 1.6Gemini Robotics-ER 1.6 improves spatial reasoning, multi-view understanding, tool use, and safety-oriented robot reasoning.
Google DeepMind2026Embodied ReasoningSafetyPlanningGenie 3: A General-Purpose World ModelGenie 3 generates interactive environments that can be explored in real time from text descriptions.
Google DeepMind2026World ModelsInteractive EnvironmentsGR00T N1: An Open Foundation Model for Generalist Humanoid RobotsGR00T N1 uses a dual-system architecture for language reasoning and continuous humanoid control.
NVIDIA2025HumanoidsVLANVIDIA Isaac GR00T-DreamsGR00T-Dreams uses world foundation models to generate synthetic robot trajectories from a single image and instruction.
NVIDIA2025Synthetic DataRobot LearningWorld ModelsGemini Robotics 1.5Gemini Robotics 1.5 turns visual observations and instructions into motor commands while supporting multi-step physical tasks.
Google DeepMind2025VLAEmbodied ReasoningRobot ControlMarble: A Multimodal World ModelMarble generates persistent 3D worlds from text, images, video, panoramas, or coarse 3D layouts.
World Labs2025Spatial IntelligenceWorld Models3DV-JEPA 2: Self-Supervised Video Models for Physical PlanningV-JEPA 2 learns predictive video representations that support visual understanding and zero-shot robot control.
Meta AI2025World ModelsVideoRobot PlanningAria Gen 2 Pilot DatasetAria Gen 2 Pilot Dataset captures daily activities with research glasses and synchronized multimodal sensor data.
Meta Reality Labs2025Egocentric DataWearablesRoboticsFigure HelixHelix is Figure's generalist humanoid VLA model for onboard perception, reasoning, and full-upper-body control.
Figure AI2025VLAHumanoidsDexterityASIMOV Benchmark for Robot Semantic SafetyASIMOV evaluates whether robot-brain foundation models understand unsafe physical situations and safety rules.
Google DeepMind2025SafetyBenchmarksEmbodied AIOpenVLA: An Open-Source Vision-Language-Action ModelOpenVLA is an open 7B-parameter VLA trained on the Open X-Embodiment dataset.
Stanford, UC Berkeley, Toyota Research Institute2024VLAOpen SourceGenie: Generative Interactive EnvironmentsGenie learns controllable interactive environments from unlabeled internet video.
Google DeepMind2024World ModelsVideoGenie 2: A Large-Scale Foundation World ModelGenie 2 generates action-controllable 3D environments from a single prompt image.
Google DeepMind2024World Models3D EnvironmentsEgo-Exo4D: First- and Third-Person Skilled Activity DatasetEgo-Exo4D pairs synchronized first-person and third-person video with language, gaze, audio, pose, and 3D signals.
Meta AI and academic partners2024Egocentric DataMultiview VideoLanguageDROID: Distributed Robot Interaction DatasetDROID provides diverse in-the-wild robot manipulation demonstrations across many scenes, tasks, and collectors.
Stanford, Berkeley, Google DeepMind and collaborators2024DatasetsManipulationRobot LearningMobile ALOHA: Low-Cost Whole-Body TeleoperationMobile ALOHA collects whole-body, bimanual mobile manipulation demonstrations with a low-cost teleoperation system.
Stanford2024TeleoperationMobile ManipulationDatasetspi0: A Vision-Language-Action Flow Model for General Robot Controlpi0 is a generalist robot policy trained on broad robot data to follow language instructions across dexterous tasks.
Physical Intelligence2024VLAGeneralist PolicyDexterityPARTNR: Planning and Reasoning Tasks in Human-Robot CollaborationPARTNR is a large benchmark for household human-robot collaboration with natural-language tasks.
Meta AI2024BenchmarksHuman-Robot CollaborationPlanningDreamerV3: Mastering Diverse Domains through World ModelsDreamerV3 uses robust normalization and objectives to learn across more than 150 tasks with one configuration.
Google DeepMind2023World ModelsGeneral AgentsRT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic ControlRT-2 co-trains vision-language models on web and robot data so semantic knowledge can influence actions.
Google DeepMind2023VLATransfer LearningOpen X-Embodiment and RT-XOpen X-Embodiment combines robot datasets across institutions and trains policies that transfer across embodiments.
Google DeepMind and 20+ institutions2023DatasetsCross-EmbodimentVLAVoyager: An Open-Ended Embodied Agent with Large Language ModelsVoyager autonomously explores Minecraft and builds a reusable library of executable skills.
NVIDIA, Caltech, UT Austin, Stanford, ASU2023Embodied AgentsPlanningGAIA-1: A Generative World Model for Autonomous DrivingGAIA-1 generates realistic driving scenarios conditioned on video, text, and vehicle actions.
Wayve2023World ModelsAutonomous DrivingRT-1: Robotics Transformer for Real-World Control at ScaleRT-1 trains one transformer policy on a large multi-task dataset of real robot demonstrations.
Google Robotics2022VLAManipulationEgo4D: Unscripted First-Person Video DatasetEgo4D is a large first-person video dataset for studying what people see, remember, manipulate, and anticipate.
Meta AI and academic partners2021Egocentric DataVideoBenchmarksDreamerV2DreamerV2 extends latent imagination with discrete representations and reaches human-level Atari performance.
Google Research2020World ModelsAtariDreamer: Reinforcement Learning with Latent ImaginationDreamer learns long-horizon behavior by propagating value gradients through imagined latent trajectories.
Google Research2019World ModelsModel-Based RLWorld ModelsA compact latent model can let an agent learn behavior inside its own predicted environment.
Google Brain / IDSIA2018World ModelsReinforcement Learning