Compare

VLA models

Compare robotic foundation models (VLA): parameters, architecture, open weights, control frequency. Sourced data.

Updated 2026-07-09 · 12 entries · JSON

Click any row to reveal details, sources and verification date.

Name Organization Country Release Params (B) Open weights Architecture Control (Hz)

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.

License : NVIDIA Open Model License (weights); Apache-2.0 (code)Modalities : vision, language, actionEmbodiments : humanoid, manipulator, cross-embodimentTraining 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

Sources : NVIDIA (GitHub) (2026-07-09)NVIDIA Research (GEAR Lab) (2025-12-15)arXiv (2025-03-18) Last verified 2026-07-09

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.

Modalities : vision, language, actionEmbodiments : manipulator, humanoid, cross-embodimentTraining 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

Sources : Google DeepMind (2025-09-25)arXiv (2025-10-03)Google AI for Developers (2026-07-09) Last verified 2026-07-09

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.

Modalities : vision, language, actionEmbodiments : manipulator, mobile, cross-embodimentTraining data : Builds on the π0.5 pretraining mixture; RECAP combines demonstrations, expert teleoperated corrections and autonomous on-robot experience with a value function

Sources : Physical Intelligence (2025-11-17)Physical Intelligence (model card) (2025-11-17)Physical Intelligence (GitHub) (2026-07-09) Last verified 2026-07-09

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.

Modalities : vision, language, actionEmbodiments : humanoidTraining data : ~500 h of teleoperated humanoid manipulation demonstrations with auto-generated natural-language annotations

Sources : Figure AI (2025-02-20)Humanoids Daily (2026-07-09) Last verified 2026-07-09

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.

License : MITModalities : vision, language, actionEmbodiments : manipulator, cross-embodimentTraining data : 970k robot manipulation episodes from the Open X-Embodiment dataset; trained on 64 A100 GPUs for 15 days

Sources : arXiv (2024-06-13)Hugging Face (model card) (2026-07-09)OpenVLA project (2026-07-09) Last verified 2026-07-09

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.

License : MITModalities : vision, language, actionEmbodiments : manipulator, cross-embodimentTraining data : 800k robot trajectories from the Open X-Embodiment dataset

Sources : arXiv (2024-05-20)Octo project (2026-07-09)GitHub (2026-07-09) Last verified 2026-07-09

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.

Modalities : vision, language, actionEmbodiments : manipulator, mobileTraining data : Internet-scale vision-language data co-fine-tuned with RT-1 robot demonstrations (collected with 13 robots over 17 months)

Sources : Google DeepMind (2023-07-28)arXiv (2023-07-28)InfoQ (2023-10-01) Last verified 2026-07-09

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.

License : Apache-2.0Modalities : vision, language, actionEmbodiments : manipulatorTraining data : Exclusively publicly available, crowd-sourced LeRobot community datasets from the Hugging Face hub (low-cost robot data)

Sources : arXiv (2025-06-02)Hugging Face (2025-06-03)Hugging Face (model card) (2026-07-09) Last verified 2026-07-09

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.

Modalities : vision, language, actionEmbodiments : mobile, legged, aerial, cross-embodimentTraining 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)

Sources : Mistral AI (2026-07-08)The Decoder (2026-07-08)Bloomberg (2026-07-08) Last verified 2026-07-09

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.

Modalities : vision, actionEmbodiments : autonomous-vehicleTraining data : Real-world driving data from geographically diverse environments (UK, US, Germany)

Sources : Wayve (2025-03-26)arXiv (2025-03-26)ADAS & Autonomous Vehicle International (2025-12-03) Last verified 2026-07-09

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.

Modalities : vision, language, action, video, sensorEmbodiments : manipulatorTraining data : Internet data plus Covariant's proprietary multimodal warehouse production data (Covariant Brain fleet: picking, sortation, induction, depalletization since 2017)

Sources : Covariant (2024-03-11)IEEE Spectrum (2024-03-11)TechCrunch (2024-08-31) Last verified 2026-07-09

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.

License : CC BY-NC-SA 4.0Modalities : vision, language, actionEmbodiments : humanoid, manipulator, cross-embodimentTraining 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

Sources : Hugging Face (model card) (2026-07-09)GlobeNewswire (AgiBot press release) (2025-03-11)arXiv (2025-03-09)OpenDriveLab (X) (2025-09-20) Last verified 2026-07-09

12 / 12 entries

D·Fairy

Bringing a physical AI solution into European industry?

D·Fairy supports robotics and embodied AI companies on their EMEA market entry: OEM and Tier 1 access, co-funded pilots, AI Act readiness.

Book an intro call