{"ok":true,"source":"tensorfeed.ai","lastUpdated":"2026-09-14","count":34,"entries":[{"id":"pi-0","name":"pi-0","org":"Physical Intelligence","category":"foundation_model","parameters":"3.3B","released":"2024-10","license":"Apache-2.0 (open weights)","paperUrl":"https://www.pi.website/blog/pi0","codeUrl":"https://github.com/Physical-Intelligence/openpi","demoUrl":"https://www.pi.website","notes":"Generalist VLA built on PaliGemma. Trained on 10K+ hours of cross-embodiment teleoperation data. The open release that broadly shifted the field toward VLA foundation models."},{"id":"pi-0.5","name":"pi-0.5","org":"Physical Intelligence","category":"foundation_model","parameters":"3.3B","released":"2025-04","license":"Apache-2.0 (open weights)","paperUrl":"https://www.pi.website/blog/pi05","codeUrl":"https://github.com/Physical-Intelligence/openpi","demoUrl":"https://www.pi.website","notes":"Successor to pi-0 with open-world generalization to homes the model never saw during training. Base checkpoint added to the openpi repo in September 2025."},{"id":"pi-0.7","name":"pi-0.7","org":"Physical Intelligence","category":"foundation_model","parameters":"5B","released":"2026-04","license":"proprietary (weights not released)","paperUrl":"https://www.pi.website/blog/pi07","codeUrl":null,"demoUrl":"https://www.pi.website","notes":"Steerable generalist VLA on a Gemma 3 4B backbone with an 860M flow-matching action expert. Conditions on strategy metadata and subgoal images, and matched Physical Intelligence's own RL-finetuned specialists on laundry, espresso and box building."},{"id":"gr00t-n1","name":"GR00T N1","org":"NVIDIA","category":"foundation_model","parameters":"2B","released":"2025-03","license":"open weights","paperUrl":"https://arxiv.org/abs/2503.14734","codeUrl":"https://huggingface.co/nvidia/GR00T-N1-2B","demoUrl":"https://developer.nvidia.com/isaac/gr00t","notes":"NVIDIA dual-system VLA targeting humanoids: System-2 reasoning over a vision-language backbone, System-1 diffusion policy. First open NVIDIA humanoid foundation model."},{"id":"gr00t-n1.7","name":"GR00T N1.7","org":"NVIDIA","category":"foundation_model","parameters":"3B","released":"2026-04","license":"open weights, commercial use permitted (NVIDIA Open Model License; code Apache-2.0)","paperUrl":"https://huggingface.co/blog/nvidia/gr00t-n1-7","codeUrl":"https://huggingface.co/nvidia/GR00T-N1.7-3B","demoUrl":"https://developer.nvidia.com/isaac/gr00t","notes":"Swaps the Eagle backbone for Cosmos-Reason2-2B (Qwen3-VL architecture) and pretrains on 20K hours of EgoScale human video using a relative end-effector action space. First GR00T release licensed for commercial deployment; shipped as early access in April 2026 before reaching general availability."},{"id":"rt-2","name":"RT-2","org":"Google DeepMind","category":"foundation_model","parameters":"55B","released":"2023-07","license":"proprietary","paperUrl":"https://robotics-transformer2.github.io","codeUrl":null,"demoUrl":null,"notes":"First VLM-to-action transformer. PaLI-X (55B) and PaLM-E backbones output robot actions as tokens. Closed but seeded the entire VLA category."},{"id":"gemini-robotics-2","name":"Gemini Robotics 2","org":"Google DeepMind","category":"foundation_model","parameters":"undisclosed","released":"2026-07","license":"proprietary (early-access partners)","paperUrl":"https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/","codeUrl":null,"demoUrl":"https://deepmind.google/models/gemini-robotics/on-device/","notes":"VLA that controls full humanoids from feet to fingertips, demonstrated on Apptronik Apollo 2, plus bi-arm robots from one checkpoint. Ships alongside Gemini Robotics On-Device 2, a local VLA that adapts to new bi-arm embodiments with typically fewer than 200 examples."},{"id":"gemini-robotics-er-2","name":"Gemini Robotics ER 2","org":"Google DeepMind","category":"foundation_model","parameters":"undisclosed","released":"2026-07","license":"proprietary (Gemini API, public preview)","paperUrl":"https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/","codeUrl":null,"demoUrl":"https://deepmind.google/models/gemini-robotics/embodied-reasoning/","notes":"Embodied reasoning VLM built on Gemini 3.5 Flash that acts as a robot's high-level brain: plans multi-step tasks, tracks progress from video, calls tools, and hands motor execution to any VLA. Adds multi-robot collaboration and is available to developers via the Gemini API and Google AI Studio."},{"id":"openvla","name":"OpenVLA","org":"Stanford / Berkeley / Toyota Research","category":"foundation_model","parameters":"7B","released":"2024-06","license":"MIT","paperUrl":"https://openvla.github.io","codeUrl":"https://github.com/openvla/openvla","demoUrl":"https://huggingface.co/openvla/openvla-7b","notes":"Open replication of RT-2 trained on Open X-Embodiment. The reference open VLA before pi-0; still widely fine-tuned for academic robotics work."},{"id":"octo","name":"Octo","org":"Berkeley AI Research","category":"foundation_model","parameters":"93M","released":"2024-05","license":"MIT","paperUrl":"https://octo-models.github.io","codeUrl":"https://github.com/octo-models/octo","demoUrl":null,"notes":"Small generalist transformer policy (93M and 27M variants) trained on 800K Open X-Embodiment trajectories. Designed as a fine-tunable base."},{"id":"rdt-1b","name":"RDT-1B","org":"Tsinghua University","category":"foundation_model","parameters":"1.2B","released":"2024-10","license":"MIT","paperUrl":"https://rdt-robotics.github.io/rdt-robotics/","codeUrl":"https://github.com/thu-ml/RoboticsDiffusionTransformer","demoUrl":null,"notes":"Largest open bimanual manipulation diffusion transformer at release. Pretrained on 1M+ multi-robot episodes, fine-tuned for bimanual ALOHA-class hardware."},{"id":"helix","name":"Helix","org":"Figure AI","category":"foundation_model","parameters":"undisclosed","released":"2025-02","license":"proprietary","paperUrl":"https://www.figure.ai/news/helix","codeUrl":null,"demoUrl":"https://www.figure.ai","notes":"Dual-system VLA launched on Figure 02 with high-rate control of the full upper body, including wrists, torso, head and individual fingers. Runs entirely onboard embedded low-power GPUs; succeeded by Helix 02 in January 2026."},{"id":"helix-02","name":"Helix 02","org":"Figure AI","category":"foundation_model","parameters":"undisclosed","released":"2026-01","license":"proprietary","paperUrl":"https://www.figure.ai/news/helix-02","codeUrl":null,"demoUrl":"https://www.figure.ai","notes":"Extends Helix to full-body loco-manipulation from pixels by adding System 0, a 10M-parameter whole-body controller running at 1 kHz trained on 1,000+ hours of human motion data. Demonstrated a four-minute autonomous dishwasher unload and reload on Figure 03."},{"id":"skild-s1","name":"S1","org":"Skild AI","category":"foundation_model","parameters":"undisclosed","released":"2026-08","license":"proprietary","paperUrl":"https://www.skild.ai/blogs/s1","codeUrl":null,"demoUrl":"https://www.skild.ai","notes":"Robotic foundation model built as an in-context learner: a single video demonstration specifies the task, with no fine-tuning or post-training. Skild reports executing unseen manipulation tasks up to 10 minutes long."},{"id":"figure-02","name":"Figure 02","org":"Figure AI","category":"humanoid","parameters":null,"released":"2024-08","license":"commercial","paperUrl":"https://www.prnewswire.com/news-releases/figure-unveils-figure-02-its-second-generation-humanoid-setting-new-standards-in-ai-and-robotics-302214889.html","codeUrl":null,"demoUrl":"https://www.figure.ai","notes":"Second-generation Figure humanoid with 16-DoF hands and 3x the onboard compute of Figure 01. Ran an 11-month deployment at BMW Plant Spartanburg, contributing to 30,000+ X3 vehicles, before Figure began retiring it in November 2025 in favor of Figure 03."},{"id":"figure-03","name":"Figure 03","org":"Figure AI","category":"humanoid","parameters":null,"released":"2025-10","license":"commercial","paperUrl":"https://www.figure.ai/news/introducing-figure-03","codeUrl":null,"demoUrl":"https://www.figure.ai/figure","notes":"Third-generation humanoid redesigned for Helix, the home and volume manufacturing at BotQ: palm cameras, fingertip tactile sensors that register about 3 grams, soft washable coverings. Arrived at BMW Plant Spartanburg in June 2026 running Helix 02."},{"id":"1x-neo","name":"1X NEO","org":"1X Technologies","category":"humanoid","parameters":null,"released":"2025-10","license":"commercial (consumer)","paperUrl":"https://www.1x.tech/discover/neo-home-robot","codeUrl":null,"demoUrl":"https://www.1x.tech","notes":"Home humanoid that 1X bills as the first consumer-ready humanoid: 30 kg (66 lb), tendon-driven joints, 22-DoF hands, $20K early access or $499/month. The Hayward factory began full-scale production in April 2026, with first US deliveries promised for 2026."},{"id":"tesla-optimus-gen3","name":"Optimus","org":"Tesla","category":"humanoid","parameters":null,"released":"2023-12","license":"commercial","paperUrl":"https://assets-ir.tesla.com/tesla-contents/IR/TSLA-Q2-2026-Update.pdf","codeUrl":null,"demoUrl":"https://www.tesla.com/AI","notes":"Gen 2 (December 2023) remains the latest body Tesla has publicly unveiled; the V3 reveal slipped from Q1 2026 and had not happened by September 2026. Tesla's Q2 2026 update says first-generation Optimus lines are being installed at Fremont, with initial builds going to its Optimus Academy for training data collection."},{"id":"apptronik-apollo","name":"Apollo 2","org":"Apptronik","category":"humanoid","parameters":null,"released":"2026-06","license":"commercial","paperUrl":"https://apptronik.com/news-collection/welcome-to-robot-park-where-apptroniks-apollo-goes-to-work","codeUrl":null,"demoUrl":"https://apptronik.com/apollo/apollo-2","notes":"Modular humanoid in bipedal and wheeled-base configurations, unveiled with Apptronik's Robot Park data-collection facility in Austin. Its fleet data trains Google DeepMind's Gemini Robotics models; Mercedes-Benz and GXO are customers, and a commercial Apollo 3 is in development."},{"id":"unitree-g1","name":"Unitree G1","org":"Unitree","category":"humanoid","parameters":null,"released":"2024-05","license":"commercial","paperUrl":"https://www.unitree.com/g1","codeUrl":"https://github.com/unitreerobotics","demoUrl":"https://www.unitree.com/g1","notes":"Compact 35 kg research humanoid (23 to 43 DoF) launched at $16K and now listed from $13.5K. Popular for academic research; NVIDIA's Isaac GR00T platform supports it. SDK and Python bindings published."},{"id":"unitree-h1","name":"Unitree H1","org":"Unitree","category":"humanoid","parameters":null,"released":"2023-08","license":"commercial","paperUrl":"https://www.unitree.com/h1","codeUrl":"https://github.com/unitreerobotics","demoUrl":"https://www.unitree.com/h1","notes":"Unitree's first full-size humanoid (about 180 cm, 47 kg), rated at 3.3 m/s, which Unitree lists as a world record. Common research platform for whole-body control papers; the 31-DoF H2 now sits above it in the lineup."},{"id":"boston-dynamics-atlas-electric","name":"Atlas (Electric)","org":"Boston Dynamics","category":"humanoid","parameters":null,"released":"2026-01","license":"commercial (enterprise)","paperUrl":"https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/","codeUrl":null,"demoUrl":"https://bostondynamics.com/products/atlas/","notes":"Product version of the all-electric Atlas, unveiled at CES 2026: 56 DoF, 1.9 m, 90 kg, 50 kg instant lift, self-swapping batteries, IP67. All 2026 units are committed to Hyundai's RMAC and Google DeepMind, which is integrating Gemini Robotics models."},{"id":"agility-digit","name":"Digit","org":"Agility Robotics","category":"humanoid","parameters":null,"released":"2023-03","license":"commercial","paperUrl":"https://www.agilityrobotics.com/content/agility-robotics-launches-next-generation-of-digit-worlds-first-human-centric-multi-purpose-robot-made-for-logistics-work","codeUrl":null,"demoUrl":"https://www.agilityrobotics.com","notes":"Bipedal humanoid built for warehouse tote handling, sold via RaaS or direct purchase, with early trials at Amazon and GXO and deployments at Schaeffler, Toyota and Mercado Libre. Agility plans first customer deliveries of Digit v5 (50 lb payload, cooperative safety) in December 2026."},{"id":"sanctuary-phoenix","name":"Phoenix","org":"Sanctuary AI","category":"humanoid","parameters":null,"released":"2023-05","license":"commercial","paperUrl":"https://sanctuary.ai/news/sanctuary-ai-unveils-phoenix-a-humanoid-general-purpose-robot-designed-for-work/","codeUrl":null,"demoUrl":"https://sanctuary.ai","notes":"Sixth-generation Sanctuary general-purpose robot running its Carbon AI control system, launched with 20-DoF hands and proprietary haptic sensing. Sanctuary now focuses on deploying its Physical AI and hydraulic hands across commercially available robot hardware."},{"id":"open-x-embodiment","name":"Open X-Embodiment","org":"Google DeepMind + 21 institutions","category":"dataset","parameters":null,"released":"2023-10","license":"CC-BY-4.0","paperUrl":"https://robotics-transformer-x.github.io","codeUrl":"https://github.com/google-deepmind/open_x_embodiment","demoUrl":"https://huggingface.co/datasets/jxu124/OpenX-Embodiment","notes":"Standardized cross-embodiment dataset spanning 22 robot embodiments and 1M+ trajectories. The pretraining substrate behind Octo, OpenVLA, and many followups."},{"id":"droid","name":"DROID","org":"Stanford / Berkeley / Toyota Research / Google","category":"dataset","parameters":null,"released":"2024-03","license":"CC-BY-4.0","paperUrl":"https://droid-dataset.github.io","codeUrl":"https://github.com/droid-dataset/droid","demoUrl":"https://huggingface.co/datasets/KarlP/droid","notes":"76K teleoperated trajectories (about 350 hours) across 564 scenes and 86 tasks, all on Franka arms, gathered by 50 collectors on three continents. Part of the open-source slice of the pi-0 pretraining mixture."},{"id":"agibot-world","name":"AgiBot World","org":"AgiBot","category":"dataset","parameters":null,"released":"2025-03","license":"CC-BY-NC-SA-4.0","paperUrl":"https://agibot-world.com","codeUrl":"https://huggingface.co/datasets/agibot-world/AgiBotWorld-Beta","demoUrl":"https://agibot-world.com","notes":"Beta release holds 1M+ trajectories (2,976 hours) from 100 robots across 100+ real-world scenarios, covering contact-rich manipulation, long-horizon planning and multi-robot collaboration. Followed the December 2024 Alpha release."},{"id":"mobile-aloha","name":"Mobile ALOHA","org":"Stanford","category":"dataset","parameters":null,"released":"2024-01","license":"MIT","paperUrl":"https://mobile-aloha.github.io","codeUrl":"https://github.com/MarkFzp/mobile-aloha","demoUrl":"https://mobile-aloha.github.io","notes":"Open hardware bimanual mobile manipulator built on a $32K budget, plus teleoperation data and training code. With 50 demonstrations per task, co-training with static ALOHA data raised success rates by up to 90%."},{"id":"bridgedata-v2","name":"BridgeData V2","org":"Berkeley AI Research","category":"dataset","parameters":null,"released":"2023-08","license":"CC-BY-4.0","paperUrl":"https://rail-berkeley.github.io/bridgedata","codeUrl":"https://github.com/rail-berkeley/bridge_data_v2","demoUrl":null,"notes":"60K+ trajectories on WidowX arms across 24 environments. Foundational for skill generalization research; included in Open X-Embodiment."},{"id":"hiw-500","name":"HIW-500 (Humanoids In-the-Wild)","org":"BitRobot / Unitree / Hugging Face","category":"dataset","parameters":null,"released":"2026-06","license":"CC-BY-4.0","paperUrl":"https://bitrobot-foundation.github.io/humanoids-in-the-wild-500-hours/","codeUrl":"https://huggingface.co/datasets/BitRobot/HIW-500-LeRobot","demoUrl":"https://huggingface.co/datasets/BitRobot/HIW-500","notes":"500+ hours and 23K+ episodes of whole-body Unitree G1 teleoperation in 12 real homes in Southeast Asia, with 148K+ subtask annotations. Published as raw ROS bag / MCAP recordings plus a LeRobot-format conversion."},{"id":"isaac-lab","name":"Isaac Lab","org":"NVIDIA","category":"simulator","parameters":null,"released":"2025-01","license":"BSD-3-Clause","paperUrl":"https://isaac-sim.github.io/IsaacLab/","codeUrl":"https://github.com/isaac-sim/IsaacLab","demoUrl":"https://developer.nvidia.com/isaac/sim","notes":"GPU-accelerated robot learning framework on Isaac Sim. Supersedes Isaac Gym. Used for sim-to-real RL training of locomotion and manipulation policies; Isaac Lab 2.0 shipped January 2025 and a 3.0 beta integrating Newton followed in March 2026."},{"id":"newton","name":"Newton","org":"Linux Foundation (NVIDIA / Google DeepMind / Disney Research)","category":"simulator","parameters":null,"released":"2026-03","license":"Apache-2.0","paperUrl":"https://developer.nvidia.com/blog/newton-adds-contact-rich-manipulation-and-locomotion-capabilities-for-industrial-robotics/","codeUrl":"https://github.com/newton-physics/newton","demoUrl":"https://newton-physics.github.io/newton/stable/","notes":"Open GPU physics engine built on NVIDIA Warp and OpenUSD with MuJoCo Warp as its primary solver, plus Disney's Kamino and deformable solvers. Reached 1.0 GA at GTC 2026 and now ships monthly releases."},{"id":"mujoco-playground","name":"MuJoCo Playground","org":"Google DeepMind","category":"simulator","parameters":null,"released":"2025-01","license":"Apache-2.0","paperUrl":"https://playground.mujoco.org","codeUrl":"https://github.com/google-deepmind/mujoco_playground","demoUrl":"https://playground.mujoco.org","notes":"JAX-based MuJoCo training suite with battery-included locomotion and manipulation environments. Designed for massively parallel sim-to-real RL."},{"id":"genesis","name":"Genesis World","org":"Genesis AI","category":"simulator","parameters":null,"released":"2026-05","license":"Apache-2.0","paperUrl":"https://genesis.ai/blog/the-role-of-simulation-in-scalable-robotics-genesis-world-10-and-the-path-forward","codeUrl":"https://github.com/Genesis-Embodied-AI/genesis-world","demoUrl":"https://genesis-world.readthedocs.io","notes":"Unified multi-physics simulator (rigid, FEM, MPM, particle) that started as the academic Genesis project in December 2024. Renamed Genesis World at 1.0, adding the Nyx path-traced renderer and Quadrants Python-to-GPU compiler, with development now backed by Genesis AI."}]}