{
  "dataset": "models",
  "updated": "2026-07-09",
  "entities": [
    {
      "id": "nvidia-isaac-gr00t-n1-7",
      "name": "NVIDIA Isaac GR00T N1.7",
      "org": "NVIDIA",
      "org_country": "US",
      "release_date": "2026-04",
      "params_b": 3,
      "architecture": "dual-system VLA: Cosmos-Reason2-2B VLM (System 2) + 16-layer diffusion-transformer action head (System 1), 40-step action horizon, relative end-effector action space",
      "open_weights": true,
      "license": "NVIDIA Open Model License (weights); Apache-2.0 (code)",
      "modalities": ["vision", "language", "action"],
      "embodiments": ["humanoid", "manipulator", "cross-embodiment"],
      "control_freq_hz": null,
      "training_data": "Several thousand hours of teleoperated demos (bimanual YAM, AgiBot Genie-1, simulated Galaxea R1 Pro, Unitree G1 locomanipulation) plus 20,000 h of EgoScale egocentric human video",
      "benchmarks": {
        "real_robot_evals": "N1.6 outperforms N1.5 on simulated manipulation benchmarks and on real YAM, AgiBot Genie-1 and Unitree G1 robots; N1.7 matches N1.6 with better generalization and language following"
      },
      "notable_en": "NVIDIA's flagship open robot foundation model line, started with GR00T N1 (first open humanoid foundation model, March 2025). N1.7 (GA April 2026) swaps in a Cosmos-Reason2-2B backbone and a relative end-effector action space shared between robot and human embodiments.",
      "notable_fr": "Ligne de modèles de fondation robotiques ouverts phare de NVIDIA, initiée avec GR00T N1 (premier modèle de fondation humanoïde ouvert, mars 2025). N1.7 (disponible en avril 2026) adopte un backbone Cosmos-Reason2-2B et un espace d'action relatif à l'effecteur, partagé entre robots et démonstrations humaines.",
      "sources": [
        {"url": "https://github.com/NVIDIA/Isaac-GR00T", "publisher": "NVIDIA (GitHub)", "date": "2026-07-09", "fields": ["params_b", "architecture", "license", "release_date", "training_data"]},
        {"url": "https://research.nvidia.com/labs/gear/gr00t-n1_6/", "publisher": "NVIDIA Research (GEAR Lab)", "date": "2025-12-15", "fields": ["architecture", "training_data", "benchmarks", "embodiments"]},
        {"url": "https://arxiv.org/abs/2503.14734", "publisher": "arXiv", "date": "2025-03-18", "fields": ["architecture", "notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "gemini-robotics-1-5",
      "name": "Gemini Robotics 1.5 (+ Gemini Robotics-ER 1.5)",
      "org": "Google DeepMind",
      "org_country": "US",
      "release_date": "2025-09",
      "params_b": null,
      "architecture": "dual-model agentic system: Gemini Robotics-ER 1.5 (embodied-reasoning orchestrator with tool calling) + Gemini Robotics 1.5 VLA executor with motion transfer and interleaved natural-language thinking",
      "open_weights": false,
      "license": null,
      "modalities": ["vision", "language", "action"],
      "embodiments": ["manipulator", "humanoid", "cross-embodiment"],
      "control_freq_hz": null,
      "training_data": "Not disclosed in detail; multi-embodiment robot data across ALOHA 2, bi-arm Franka and Apptronik Apollo, leveraged via a Motion Transfer mechanism on top of Gemini multimodal pretraining",
      "benchmarks": {
        "embodied_reasoning": "ER 1.5 achieves the highest aggregate performance on 15 academic embodied-reasoning benchmarks"
      },
      "notable_en": "Google DeepMind's agentic robotics stack: ER 1.5 plans, reasons spatially and calls tools (available via the Gemini API; ER 1.6 followed in 2026), while the Gemini Robotics 1.5 VLA executes and transfers motions across embodiments. Model weights remain closed, with the VLA limited to selected partners.",
      "notable_fr": "Pile robotique agentique de Google DeepMind : ER 1.5 planifie, raisonne spatialement et appelle des outils (disponible via l'API Gemini ; ER 1.6 a suivi en 2026), tandis que le VLA Gemini Robotics 1.5 exécute et transfère les mouvements entre robots. Les poids restent fermés, le VLA étant réservé à des partenaires sélectionnés.",
      "sources": [
        {"url": "https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/", "publisher": "Google DeepMind", "date": "2025-09-25", "fields": ["release_date", "architecture", "notable_en"]},
        {"url": "https://arxiv.org/abs/2510.03342", "publisher": "arXiv", "date": "2025-10-03", "fields": ["architecture", "training_data", "benchmarks", "embodiments"]},
        {"url": "https://ai.google.dev/gemini-api/docs/robotics-overview", "publisher": "Google AI for Developers", "date": "2026-07-09", "fields": ["notable_en", "open_weights"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "pi-star-0-6",
      "name": "π*0.6 (Physical Intelligence)",
      "org": "Physical Intelligence",
      "org_country": "US",
      "release_date": "2025-11",
      "params_b": null,
      "architecture": "VLA built on a 5B-parameter VLM augmented with an action expert; RL post-training via RECAP (RL with Experience and Corrections via Advantage-conditioned Policies)",
      "open_weights": false,
      "license": null,
      "modalities": ["vision", "language", "action"],
      "embodiments": ["manipulator", "mobile", "cross-embodiment"],
      "control_freq_hz": null,
      "training_data": "Builds on the π0.5 pretraining mixture; RECAP combines demonstrations, expert teleoperated corrections and autonomous on-robot experience with a value function",
      "benchmarks": {
        "recap_gains": "RECAP roughly doubles throughput and more than halves failure rates versus the π0.6 base model on espresso making, laundry folding and box assembly"
      },
      "notable_en": "Latest VLA from the best-funded US robot foundation model startup, trained with RECAP to learn from its own experience; ran espresso making, laundry folding and factory box packing for hours uninterrupted. Unlike π0/π0.5 (open-sourced via openpi), π*0.6 weights had not been released as of mid-2026.",
      "notable_fr": "Dernier VLA de la startup américaine la mieux financée sur les modèles de fondation robotiques, entraîné avec RECAP pour apprendre de sa propre expérience ; il a préparé des cafés, plié du linge inconnu et assemblé des cartons en usine pendant des heures sans interruption. Contrairement à π0/π0.5 (ouverts via openpi), les poids de π*0.6 n'étaient pas publiés mi-2026.",
      "sources": [
        {"url": "https://www.pi.website/blog/pistar06", "publisher": "Physical Intelligence", "date": "2025-11-17", "fields": ["release_date", "architecture", "benchmarks", "notable_en"]},
        {"url": "https://website.pi-asset.com/pi06star/PI06_model_card.pdf", "publisher": "Physical Intelligence (model card)", "date": "2025-11-17", "fields": ["architecture", "training_data"]},
        {"url": "https://github.com/Physical-Intelligence/openpi", "publisher": "Physical Intelligence (GitHub)", "date": "2026-07-09", "fields": ["open_weights", "notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "figure-helix",
      "name": "Helix",
      "org": "Figure AI",
      "org_country": "US",
      "release_date": "2025-02",
      "params_b": 7,
      "architecture": "dual-system VLA (System 1/System 2): 7B internet-pretrained VLM at 7-9 Hz + 80M reactive visuomotor policy at 200 Hz",
      "open_weights": false,
      "license": null,
      "modalities": ["vision", "language", "action"],
      "embodiments": ["humanoid"],
      "control_freq_hz": 200,
      "training_data": "~500 h of teleoperated humanoid manipulation demonstrations with auto-generated natural-language annotations",
      "benchmarks": {},
      "notable_en": "First VLA to output continuous high-rate control of the entire humanoid upper body (wrists, individual fingers, torso, head) and to run fully onboard embedded low-power GPUs, including zero-shot two-robot collaboration with a single set of weights.",
      "notable_fr": "Premier VLA à contrôler en continu et à haute fréquence tout le haut du corps d'un humanoïde (poignets, doigts, torse, tête) et à fonctionner entièrement embarqué sur GPU basse consommation, y compris une collaboration à deux robots en zero-shot avec un seul jeu de poids.",
      "sources": [
        {"url": "https://www.figure.ai/news/helix", "publisher": "Figure AI", "date": "2025-02-20", "fields": ["release_date", "params_b", "architecture", "control_freq_hz", "training_data", "notable_en"]},
        {"url": "https://www.humanoidsdaily.com/news/figure-ai-reorganizes-to-boost-humanoid-learning-with-new-helix-ai-model", "publisher": "Humanoids Daily", "date": "2026-07-09", "fields": ["notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "openvla",
      "name": "OpenVLA",
      "org": "Stanford / UC Berkeley / TRI / Google DeepMind (academic consortium)",
      "org_country": "US",
      "release_date": "2024-06",
      "params_b": 7,
      "architecture": "autoregressive VLA: Prismatic VLM (Llama-2-7B language backbone + fused SigLIP and DINOv2 vision encoders), actions as discrete tokens",
      "open_weights": true,
      "license": "MIT",
      "modalities": ["vision", "language", "action"],
      "embodiments": ["manipulator", "cross-embodiment"],
      "control_freq_hz": null,
      "training_data": "970k robot manipulation episodes from the Open X-Embodiment dataset; trained on 64 A100 GPUs for 15 days",
      "benchmarks": {
        "cross_embodiment_evals": "outperforms RT-2-X (55B) by 16.5 points absolute success rate across 29 tasks despite 7x fewer parameters"
      },
      "notable_en": "The reference open-source VLA: beat the closed 55B RT-2-X with a 7B model and released all checkpoints and training code under MIT, catalyzing the academic and startup VLA ecosystem.",
      "notable_fr": "Le VLA open source de référence : il a surpassé le modèle fermé RT-2-X (55 Md) avec seulement 7 Md de paramètres et publié tous les poids et le code d'entraînement sous licence MIT, catalysant l'écosystème VLA académique et startup.",
      "sources": [
        {"url": "https://arxiv.org/abs/2406.09246", "publisher": "arXiv", "date": "2024-06-13", "fields": ["release_date", "params_b", "architecture", "training_data", "benchmarks"]},
        {"url": "https://huggingface.co/openvla/openvla-7b", "publisher": "Hugging Face (model card)", "date": "2026-07-09", "fields": ["license", "open_weights", "training_data"]},
        {"url": "https://openvla.github.io/", "publisher": "OpenVLA project", "date": "2026-07-09", "fields": ["architecture", "notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "octo",
      "name": "Octo",
      "org": "UC Berkeley / Stanford / CMU / Google DeepMind (academic consortium)",
      "org_country": "US",
      "release_date": "2024-05",
      "params_b": 0.093,
      "architecture": "transformer policy with diffusion action head; flexible observation and action spaces (Octo-Small 27M, Octo-Base 93M)",
      "open_weights": true,
      "license": "MIT",
      "modalities": ["vision", "language", "action"],
      "embodiments": ["manipulator", "cross-embodiment"],
      "control_freq_hz": null,
      "training_data": "800k robot trajectories from the Open X-Embodiment dataset",
      "benchmarks": {
        "multi_robot_evals": "evaluated on 9 robot platforms across 4 institutions, with efficient fine-tuning to new observation and action spaces"
      },
      "notable_en": "First fully open-source generalist robot policy (data, weights and training pipeline); its compact diffusion-head transformer became a standard academic baseline for cross-embodiment manipulation.",
      "notable_fr": "Première politique robotique généraliste entièrement open source (données, poids et pipeline d'entraînement) ; son transformer compact à tête de diffusion est devenu une référence académique pour la manipulation multi-robots.",
      "sources": [
        {"url": "https://arxiv.org/abs/2405.12213", "publisher": "arXiv", "date": "2024-05-20", "fields": ["release_date", "params_b", "architecture", "training_data", "benchmarks"]},
        {"url": "https://octo-models.github.io/", "publisher": "Octo project", "date": "2026-07-09", "fields": ["architecture", "notable_en"]},
        {"url": "https://github.com/octo-models/octo", "publisher": "GitHub", "date": "2026-07-09", "fields": ["license", "open_weights"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "rt-2",
      "name": "RT-2 (Robotic Transformer 2)",
      "org": "Google DeepMind",
      "org_country": "US",
      "release_date": "2023-07",
      "params_b": 55,
      "architecture": "VLA co-fine-tuned from web-scale VLMs (PaLI-X 55B and PaLM-E 12B variants); robot actions emitted as text tokens",
      "open_weights": false,
      "license": null,
      "modalities": ["vision", "language", "action"],
      "embodiments": ["manipulator", "mobile"],
      "control_freq_hz": null,
      "training_data": "Internet-scale vision-language data co-fine-tuned with RT-1 robot demonstrations (collected with 13 robots over 17 months)",
      "benchmarks": {
        "emergent_skills": "up to 3x improvement over RT-1 on emergent skill evaluations; success rate on unseen tasks roughly doubled (~32% to ~62%)"
      },
      "notable_en": "The founding VLA: first to show that web-scale vision-language knowledge transfers directly to robot control by treating actions as text tokens. It spawned the RT-X / Open X-Embodiment cross-lab effort that underpins most later open VLAs. Research artifact, never released as weights or product.",
      "notable_fr": "Le VLA fondateur : premier à démontrer que les connaissances vision-langage à l'échelle du web se transfèrent directement au contrôle robotique en traitant les actions comme des tokens de texte. Il a engendré l'initiative RT-X / Open X-Embodiment sur laquelle reposent la plupart des VLA ouverts ultérieurs. Artefact de recherche, jamais publié en poids ni en produit.",
      "sources": [
        {"url": "https://deepmind.google/blog/rt-2-new-model-translates-vision-and-language-into-action/", "publisher": "Google DeepMind", "date": "2023-07-28", "fields": ["release_date", "architecture", "benchmarks", "training_data"]},
        {"url": "https://arxiv.org/abs/2307.15818", "publisher": "arXiv", "date": "2023-07-28", "fields": ["params_b", "architecture", "benchmarks"]},
        {"url": "https://www.infoq.com/news/2023/10/deepmind-robot-transformer/", "publisher": "InfoQ", "date": "2023-10-01", "fields": ["params_b", "notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "smolvla",
      "name": "SmolVLA",
      "org": "Hugging Face (LeRobot team)",
      "org_country": "US",
      "release_date": "2025-06",
      "params_b": 0.45,
      "architecture": "compact VLA: SmolVLM-2 backbone + flow-matching action expert conditioned on multi-camera views, robot state and language; asynchronous inference",
      "open_weights": true,
      "license": "Apache-2.0",
      "modalities": ["vision", "language", "action"],
      "embodiments": ["manipulator"],
      "control_freq_hz": null,
      "training_data": "Exclusively publicly available, crowd-sourced LeRobot community datasets from the Hugging Face hub (low-cost robot data)",
      "benchmarks": {
        "real_world_and_sim": "~78% average success rate on real-world task suites; competitive with much larger VLAs on LIBERO and Meta-World"
      },
      "notable_en": "A 450M open VLA trained only on crowd-sourced community data, small enough to train on a single consumer GPU and run on CPU or a MacBook; asynchronous inference yields ~30% faster response and ~2x task throughput. Built by Hugging Face's Paris-based LeRobot team.",
      "notable_fr": "Un VLA ouvert de 450 M de paramètres entraîné uniquement sur des données communautaires, assez petit pour être entraîné sur un seul GPU grand public et tourner sur CPU ou MacBook ; l'inférence asynchrone apporte ~30 % de latence en moins et ~2x de débit de tâches. Développé par l'équipe LeRobot de Hugging Face, basée à Paris.",
      "sources": [
        {"url": "https://arxiv.org/abs/2506.01844", "publisher": "arXiv", "date": "2025-06-02", "fields": ["release_date", "params_b", "architecture", "training_data", "benchmarks"]},
        {"url": "https://huggingface.co/blog/smolvla", "publisher": "Hugging Face", "date": "2025-06-03", "fields": ["params_b", "training_data", "benchmarks", "notable_en"]},
        {"url": "https://huggingface.co/lerobot/smolvla_base", "publisher": "Hugging Face (model card)", "date": "2026-07-09", "fields": ["license", "open_weights"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "mistral-robostral-navigate",
      "name": "Robostral Navigate",
      "org": "Mistral AI",
      "org_country": "FR",
      "release_date": "2026-07",
      "params_b": 8,
      "architecture": "pointing-based navigation model: grounding-specialized Mistral VLM predicts target image coordinates and desired orientation in the current camera view, with local displacement fallback",
      "open_weights": false,
      "license": null,
      "modalities": ["vision", "language", "action"],
      "embodiments": ["mobile", "legged", "aerial", "cross-embodiment"],
      "control_freq_hz": null,
      "training_data": "~400,000 simulated trajectories across 6,000 scenes, generated entirely in-house without open-source VLMs; further improved with online reinforcement learning (+3.2 pts reported)",
      "benchmarks": {
        "R2R-CE": "76.6% success on validation unseen (79.4% on validation seen); beats the best single-camera method by 9.7 pts and the best depth/multi-camera system by 4.5 pts"
      },
      "notable_en": "Mistral AI's entry into physical AI (announced 8 July 2026): an 8B model steering wheeled, legged and flying robots from a single RGB camera and plain-language instructions, trained entirely in simulation. Availability and licensing had not been announced at launch.",
      "notable_fr": "L'entrée de Mistral AI dans l'IA physique (annoncée le 8 juillet 2026) : un modèle de 8 Md de paramètres qui pilote des robots à roues, à pattes et volants à partir d'une seule caméra RGB et d'instructions en langage naturel, entraîné entièrement en simulation. Disponibilité et licence non annoncées au lancement.",
      "sources": [
        {"url": "https://mistral.ai/news/robostral-navigate/", "publisher": "Mistral AI", "date": "2026-07-08", "fields": ["params_b", "architecture", "training_data", "benchmarks", "embodiments"]},
        {"url": "https://the-decoder.com/mistral-enters-robotics-with-robostral-navigate-an-8b-model-that-steers-robots-using-just-one-camera/", "publisher": "The Decoder", "date": "2026-07-08", "fields": ["release_date", "params_b", "training_data", "open_weights"]},
        {"url": "https://www.bloomberg.com/news/articles/2026-07-08/mistral-ai-releases-robotics-model-to-support-physical-ai-push", "publisher": "Bloomberg", "date": "2026-07-08", "fields": ["release_date", "notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "wayve-gaia-2",
      "name": "GAIA-2",
      "org": "Wayve",
      "org_country": "GB",
      "release_date": "2025-03",
      "params_b": null,
      "architecture": "latent-diffusion multi-camera generative world model conditioned on ego-vehicle dynamics, agent configurations, environmental factors and road semantics",
      "open_weights": false,
      "license": null,
      "modalities": ["vision", "action"],
      "embodiments": ["autonomous-vehicle"],
      "control_freq_hz": null,
      "training_data": "Real-world driving data from geographically diverse environments (UK, US, Germany)",
      "benchmarks": {},
      "notable_en": "Europe's flagship driving world model: generates controllable, spatiotemporally consistent multi-camera video, including rare and safety-critical scenarios for training and validating Wayve's driving AI. Succeeded in December 2025 by GAIA-3 (15B parameters, double GAIA-2's size), offered for AV validation.",
      "notable_fr": "Le modèle du monde de conduite phare en Europe : il génère des vidéos multi-caméras contrôlables et cohérentes dans l'espace-temps, y compris des scénarios rares et critiques pour entraîner et valider l'IA de conduite de Wayve. Remplacé en décembre 2025 par GAIA-3 (15 Md de paramètres, le double de GAIA-2), proposé pour la validation de conduite autonome.",
      "sources": [
        {"url": "https://wayve.ai/press/wayve-unveils-gaia2/", "publisher": "Wayve", "date": "2025-03-26", "fields": ["release_date", "architecture", "training_data", "notable_en"]},
        {"url": "https://arxiv.org/abs/2503.20523", "publisher": "arXiv", "date": "2025-03-26", "fields": ["architecture", "modalities"]},
        {"url": "https://www.autonomousvehicleinternational.com/news/ai-sensor-fusion/wayves-gaia-3-generative-world-model-now-available-for-autonomous-driving-validation.html", "publisher": "ADAS & Autonomous Vehicle International", "date": "2025-12-03", "fields": ["notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "covariant-rfm-1",
      "name": "RFM-1",
      "org": "Covariant",
      "org_country": "US",
      "release_date": "2024-03",
      "params_b": 8,
      "architecture": "8B any-to-any multimodal transformer performing autoregressive next-token prediction across text, images, video, robot actions and numerical sensor readings",
      "open_weights": false,
      "license": null,
      "modalities": ["vision", "language", "action", "video", "sensor"],
      "embodiments": ["manipulator"],
      "control_freq_hz": null,
      "training_data": "Internet data plus Covariant's proprietary multimodal warehouse production data (Covariant Brain fleet: picking, sortation, induction, depalletization since 2017)",
      "benchmarks": {},
      "notable_en": "Early commercial robotics foundation model for warehouse manipulation, including a learned physics world model that predicts video outcomes of candidate actions. In August 2024 Amazon hired Covariant's founders and about a quarter of its staff and took a non-exclusive license to its models.",
      "notable_fr": "Modèle de fondation robotique commercial pionnier pour la manipulation en entrepôt, doté d'un modèle du monde physique appris qui prédit en vidéo l'issue des actions candidates. En août 2024, Amazon a recruté les fondateurs de Covariant et environ un quart de l'équipe, avec une licence non exclusive sur ses modèles.",
      "sources": [
        {"url": "https://covariant.ai/insights/introducing-rfm-1-giving-robots-human-like-reasoning-capabilities/", "publisher": "Covariant", "date": "2024-03-11", "fields": ["release_date", "params_b", "architecture", "training_data"]},
        {"url": "https://spectrum.ieee.org/covariant-foundation-model", "publisher": "IEEE Spectrum", "date": "2024-03-11", "fields": ["params_b", "architecture", "notable_en"]},
        {"url": "https://techcrunch.com/2024/08/31/amazon-hires-the-founders-of-robotics-ai-startup-covariant/", "publisher": "TechCrunch", "date": "2024-08-31", "fields": ["notable_en"]}
      ],
      "last_verified": "2026-07-09"
    },
    {
      "id": "agibot-go-1",
      "name": "AgiBot GO-1 (Genie Operator-1)",
      "org": "AgiBot (Zhiyuan Robotics) / OpenDriveLab",
      "org_country": "CN",
      "release_date": "2025-03",
      "params_b": 3,
      "architecture": "ViLLA (Vision-Language-Latent-Action): InternVL2.5-2B VLM backbone + MoE with a 24-layer latent planner predicting latent action tokens and a high-frequency action expert",
      "open_weights": true,
      "license": "CC BY-NC-SA 4.0",
      "modalities": ["vision", "language", "action"],
      "embodiments": ["humanoid", "manipulator", "cross-embodiment"],
      "control_freq_hz": null,
      "training_data": "AgiBot World dataset: over 1 million real-robot trajectories across 217 tasks in 5 application domains, plus cross-embodiment and human video data for the latent planner",
      "benchmarks": {
        "agibot_world_tasks": "reported +32% success rate over prior state-of-the-art generalist policies on real-world complex tasks"
      },
      "notable_en": "China's flagship open embodied foundation model, introducing the ViLLA framework where latent action tokens bridge vision-language inputs and robot control. Announced March 2025 and fully open-sourced (weights on Hugging Face) in September 2025, alongside the largest open real-robot dataset.",
      "notable_fr": "Modèle de fondation incarné ouvert phare de la Chine, qui introduit le cadre ViLLA où des tokens d'action latents relient les entrées vision-langage au contrôle du robot. Annoncé en mars 2025 et intégralement open source (poids sur Hugging Face) en septembre 2025, aux côtés du plus grand jeu de données robotique réel ouvert.",
      "sources": [
        {"url": "https://huggingface.co/agibot-world/GO-1", "publisher": "Hugging Face (model card)", "date": "2026-07-09", "fields": ["params_b", "license", "architecture", "training_data", "open_weights"]},
        {"url": "https://www.globenewswire.com/news-release/2025/03/11/3040608/0/en/AgiBot-GO-1-The-Evolution-of-Generalist-Embodied-Foundation-Model-from-VLA-to-ViLLA.html", "publisher": "GlobeNewswire (AgiBot press release)", "date": "2025-03-11", "fields": ["release_date", "architecture", "benchmarks"]},
        {"url": "https://arxiv.org/abs/2503.06669", "publisher": "arXiv", "date": "2025-03-09", "fields": ["training_data", "architecture"]},
        {"url": "https://x.com/OpenDriveLab/status/1969243929901220110", "publisher": "OpenDriveLab (X)", "date": "2025-09-20", "fields": ["open_weights", "notable_en"]}
      ],
      "last_verified": "2026-07-09"
    }
  ]
}