Robot AI Model Comparison

Robot foundation models compared: developer, release date, parameters, open weights, license and supported robots for Gemini Robotics, NVIDIA GR00T, Physical Intelligence π, Figure Helix and more.

Updated November 16, 202618 modelsEvery row sourcedMethodologyCite or embed

Models tracked
18
from 8 developers
With open weights
8
fully or partly
Newest
Helix 2.5
2026-09-17, figure.ai (opens in a new tab)
Released in 2026
5
new models or major versions

Overview

Models in our list by release year

Download chart (SVG)

Compare two models

MeasureGemini Robotics 2Isaac GR00T N1.7
Developer
Released
Parameters
Weights
License
Architecture
Robots
Source

All models

18 rows

ModelDeveloperReleasedParametersOpen weightsLicenseArchitectureRobotsSource
Helix 2.5Figure AINot disclosedNoProprietaryPretrained on Index human-behaviour dataset; 56% zero-shot success in 30 unseen homesFigure humanoidsfigure.ai (opens in a new tab)
Gemini Robotics 2Google DeepMindNot disclosedNoProprietaryVLA, ER 2 reasoning model and On-Device 2Whole-body humanoids (Apptronik Apollo 2), bi-arm robotsdeepmind.google (opens in a new tab)
Isaac GR00T N1.7NVIDIA3BYesNVIDIA Open Model LicenseCosmos-Reason2-2B VLM plus 32-layer DiT, pretrained on 20,854 hours of human egocentric videoUnitree G1, YAM, AGIBot Genie 1huggingface.co (opens in a new tab)
π0.7Physical IntelligenceNot disclosedNoProprietarySteerable VLA trained with multimodal prompts and visual subgoalsMultiple platforms incl. UR5epi.website (opens in a new tab)
Helix 02Figure AINot disclosedNoProprietaryS0 whole-body controller, S1, S2; trained on 1,000+ hours of human motionFigure 03figure.ai (opens in a new tab)
Isaac GR00T N1.6NVIDIANot stated on pageYesNVIDIA Open Model LicenseCosmos-2B VLM variant plus 32-layer DiTBimanual YAM, Agibot Genie-1, Unitree G1research.nvidia.com (opens in a new tab)
Gemini Robotics 1.5Google DeepMindNot disclosedNoProprietaryVLA with thinking and motion transfer, plus ER 1.5 reasoning modelALOHA, Franka, Apolloarxiv.org (opens in a new tab)
Gemini Robotics On-DeviceGoogle DeepMindNot disclosedNoProprietaryOn-device VLABi-arm robots, Franka FR3, Apollodeepmind.google (opens in a new tab)
SmolVLAHugging Face450MYesApache 2.0Compact VLA trained on community LeRobot dataLow-cost arms (SO-100)huggingface.co (opens in a new tab)
π0.5Physical IntelligenceAction expert 300M on the π0 baseYesApache 2.0Co-trained VLA with open-world generalizationMobile manipulatorspi.website (opens in a new tab)
Isaac GR00T N1NVIDIANot stated on pageYesNVIDIA Open Model LicenseDual system: VLM plus diffusion transformer action headHumanoids, cross-embodimentnvidianews.nvidia.com (opens in a new tab)
Gemini RoboticsGoogle DeepMindNot disclosedNoProprietaryVLA built on Gemini 2.0ALOHA 2, Franka, Apptronik Apollodeepmind.google (opens in a new tab)
HelixFigure AIS2: 7B VLM; S1: 80MNoProprietarySystem 2 VLM plus System 1 visuomotor policyFigure humanoidsfigure.ai (opens in a new tab)
π0Physical Intelligence3B VLM plus action expertYesApache 2.0VLM (PaliGemma) plus flow-matching action expert8 robot typespi.website (opens in a new tab)
OpenVLAStanford and partners7BYesMITLlama 2 based VLA trained on 970k episodesOpen X-Embodiment robotsopenvla.github.io (opens in a new tab)
OctoUC Berkeley, Stanford, CMU27M and 93MYesMITTransformer diffusion policy trained on 800k trajectoriesOpen X-Embodiment robotsocto-models.github.io (opens in a new tab)
RT-2Google DeepMindUp to 55BNoProprietaryVLA built on PaLI-X and PaLM-EEveryday Robots armdeepmind.google (opens in a new tab)
Redwood AI1Xn/a160MNoProprietaryOn-board VLA for home humanoid1X NEO1x.tech (opens in a new tab)

Methodology

One row per model or major version, taken from the developer's own announcement, model card or paper. We record parameters, weights and license only as stated there; when a source does not state a value we say so.

Limits. We do not compare benchmark scores because developers test on different tasks and robots. Closed models rarely disclose size, so blanks are expected.

Data note. Facts from each developer's own announcement, model card or paper. Parameters are listed only where the source states them. Benchmarks are not comparable across developers and are left out on purpose.

All 18 sources
  1. https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization (opens in a new tab)
  2. https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/ (opens in a new tab)
  3. https://huggingface.co/blog/nvidia/gr00t-n1-7 (opens in a new tab)
  4. https://www.pi.website/blog/pi07 (opens in a new tab)
  5. https://www.figure.ai/news/helix-02 (opens in a new tab)
  6. https://research.nvidia.com/labs/gear/gr00t-n1_6/index.html (opens in a new tab)
  7. https://arxiv.org/html/2510.03342v3 (opens in a new tab)
  8. https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/ (opens in a new tab)
  9. https://huggingface.co/blog/smolvla (opens in a new tab)
  10. https://www.pi.website/blog/pi05 (opens in a new tab)
  11. https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks (opens in a new tab)
  12. https://deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world/ (opens in a new tab)
  13. https://www.figure.ai/news/helix (opens in a new tab)
  14. https://www.pi.website/blog/pi0 (opens in a new tab)
  15. https://openvla.github.io/ (opens in a new tab)
  16. https://octo-models.github.io/ (opens in a new tab)
  17. https://deepmind.google/discover/blog/rt-2-new-model-translates-vision-and-language-into-action/ (opens in a new tab)
  18. https://www.1x.tech/discover/redwood-ai (opens in a new tab)

Spotted an error or a missing row? Write to the newsroom via Contact. We correct and date every change.

How to cite or embed

Free to quote, chart and embed. Please credit i10X News and link to this page.

Download CSVJSON

More from the data desk

All datasets