Skip to content
embodied.guide

Back to the AI models

Shanghai AI LaboratoryChina2025

EO-1

An AI model: software that a robot uses to perceive, plan or move. This page describes the software, not a robot.

RobotsG1by Agibot B

Visit website

eo-robotics.ai

Key figures as stated by the developer

6 of 6 stated
  • Model type ⁠A vision-language-action embodied reasoning
  • Controls ⁠A arms and hands
  • Inputs ⁠A camera language robot state
  • Access ⁠C open download
  • Parameters ⁠A 3,000,000,000
  • Base model ⁠A Qwen2.5-VL-3B-Instruct
How to read the figures
50HzA
The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.

What the data labels mean

Versions and notes in the key figures (6)

Model type

Model card: "EO-1 model is a Vision-Language-Action (VLA) model that adopts a single unified decoder-only transformer"; it is built "for multimodal embodied reasoning and robot control, enabling seamless perception, planning, reasoning, and acting in single model".

Controls

Model card, Output: "Output Type(s): Actions, Language". The real-robot tests are arm and dual-arm manipulation tasks: "Agibot G-1 Long-Horizon Dexterity (4 Tasks)."

Inputs

Model card, Input Type: "Vision: Image Frames, Video"; "State: Robot Proprioception"; "Language Instruction: Text, Pointing, Bounding Box, etc.". Image frames and video = camera, robot proprioception = robot state.

Access

Project page: "EO-1 is released with full openness, including model weights, training code, and all components of the interleaved embodied dataset".

Parameters

Model card: "an open-source unified embodied foundation model comprising 3B parameters".

Base model

Model card: the tokens are processed by the "shared transformer backbone, whose weights are initialized from Qwen2.5-VL"; base model listed as "Qwen/Qwen2.5-VL-3B-Instruct".

Full specifications

10 of 13 fields stated
The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field Value Source and caveats
Input and output
Model typevision-language-action embodied reasoningA

Model card: "EO-1 model is a Vision-Language-Action (VLA) model that adopts a single unified decoder-only transformer"; it is built "for multimodal embodied reasoning and robot control, enabling seamless perception, planning, reasoning, and acting in single model".

Inputscamera language robot stateA

Model card, Input Type: "Vision: Image Frames, Video"; "State: Robot Proprioception"; "Language Instruction: Text, Pointing, Bounding Box, etc.". Image frames and video = camera, robot proprioception = robot state.

Controlsarms and handsA

Model card, Output: "Output Type(s): Actions, Language". The real-robot tests are arm and dual-arm manipulation tasks: "Agibot G-1 Long-Horizon Dexterity (4 Tasks)."
Size and speed
Parameters3,000,000,000A

Model card: "an open-source unified embodied foundation model comprising 3B parameters".

Control ratenot stated

Note Paper: the model denoises "a 16-step action chunk through 10 denoising iterations" for execution. ⁠B
Where it runs and what it is built on
On robot or cloudnot stated

Note Model card: inference "requires ~6.5GB GPU memory". ⁠A

Base modelQwen2.5-VL-3B-InstructA

Model card: the tokens are processed by the "shared transformer backbone, whose weights are initialized from Qwen2.5-VL"; base model listed as "Qwen/Qwen2.5-VL-3B-Instruct".
Openness
Accessopen downloadC

Project page: "EO-1 is released with full openness, including model weights, training code, and all components of the interleaved embodied dataset".

LicenseMITA

Model card header: "license: mit".

Free or paidfully open sourceC

Project page: "EO-1 is released with full openness, including model weights, training code, and all components of the interleaved embodied dataset"; the model card gives "license: mit".
The developer’s own claims
Learns new tasksneeds retrainingB

Developer claim: the long-horizon AgiBot G-1 tasks are learned by fine-tuning: "Specializing fine-tuning is performed using 150 demonstration trajectories per task".

Works across robotsyesB

Developer claim: "We evaluate its performance on both short-horizon and long-horizon tasks using different robots, including Franka Panda, WidowX 250 S, Agibot G-1, and Lerobot SO100."

Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Not stated: Degrees of freedom controlled

Notes

  • The EO Robotics team at Shanghai AI Laboratory describes EO-1 as an open-source unified embodied foundation model with 3B parameters.
  • The EO-1 paper was first submitted to arXiv on 28 August 2025.
  • EO Robotics is part of the Intelligent Photonics and Electronics Center of Shanghai AI Laboratory in Xuhui, Shanghai.
  • The EO Robotics website is copyrighted by Shanghai Artificial Intelligence Laboratory.

Sources

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.