What Is EO-1?

EO-1 is introduced in the paper now titled EO-1: Interleaved Vision-Text-Action Pretraining for General Robot Control (arXiv 2508.21112) by Delin Qu and fourteen co-authors, first submitted on 28 August 2025 and last revised on 15 October 2025 (v4). The official project page uses the subtitle An Open Unified Embodied Foundation Model for General Robot Control. The abstract introduces EO-Robotics, made of the EO-1 model and the EO-Data1.5M dataset.

EO-1 is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. What the site states is custom egocentric collection of real human demonstrations, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. The only language data the site states is one task instruction per episode.

The abstract says recent VLA models, co-trained on large-scale robot and visual-text data, still fall short of human-level flexibility in interleaved reasoning and interaction. EO-1 targets that gap by pretraining on data where vision, text and action are interleaved, so one model both reasons about a scene and acts in it.

Per the Hugging Face model card, EO-1 is a vision-language-action model with one decoder-only transformer, a discrete language-modeling head for embodied reasoning and a continuous flow-matching head for robot action generation. Language instruction, image observations, robot state and noisy action become one interleaved token sequence, and the backbone weights are initialized from Qwen2.5-VL.

The abstract says it is trained through synergies between auto-regressive decoding and flow matching denoising. The README describes an open-source model of 3B parameters. Listed inputs are image frames or video, robot proprioception and a language instruction; outputs are actions and text.

The abstract describes EO-Data1.5M as a massive, high-quality multimodal embodied reasoning dataset of over 1.5 million samples, emphasizing interleaved vision-text-action comprehension. The project page splits the pretraining corpus into web multimodal data, robot control data and interleaved embodied data.

The README names robot control sources such as AgiBotWorld, Open X-Embodiment, RoboMIND and SO100-Community. The project page says model weights, training code and all components of the interleaved dataset are released.

The project page says EO-1 surpasses existing open-source models on ERQA, LIBERO, SimplerEnv and the authors’ own EO-Bench, and reports real-robot tests. The README lists Franka Panda, WidowX 250 S, AgiBot G-1 and LeRobot SO100 platforms. Named long-horizon tasks are making a breakfast sandwich, roasting a beef steak, folding household clothes and sorting grocery items.

These are the authors’ claims on their own setups. Firsthand does not reproduce them and quotes no scores.

The site’s Interleave-VLA entry covers a framework whose instructions mix text and images, such as crops or sketches. EO-1 instead interleaves vision, text and action in the pretraining data. Both are third-party research.

EO-1 describes itself as a unified embodied foundation model, the idea the site defines in its robot foundation model and vision-language-action model entries. Firsthand ranks none of these.

This entry draws on the arXiv abstract page, the official project page, the official GitHub README and the Hugging Face model and paper pages. The full paper text was not read, and the licence is Not stated here. The sources name no human egocentric video, stereo depth, hand pose or IMU data, so how those would suit EO-1 is Not stated.

When a program needs real human egocentric demonstrations of tasks, environments or objects that public data does not cover, specified in writing. Firsthand does not build or train models, and the language it delivers is one task instruction per episode.

Confirm the spec on a Spec Pilot, a small fast first batch. Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour.

What is EO-1?

A published open-source unified embodied foundation model (arXiv 2508.21112) that handles image, text, video and action inputs in one decoder-only transformer for reasoning and robot control.

What is EO-Data1.5M?

The paired pretraining dataset: over 1.5 million samples emphasizing interleaved vision-text-action comprehension, per the abstract.

Is EO-1 a VLA model?

Its model card calls it a vision-language-action model that also produces text for embodied reasoning, through a language head and a flow-matching action head.

Is the code or model open?

The project page says model weights, training code and dataset components are released. The licence is Not stated in the sources read.

Does Firsthand data train EO-1?

No claim is made. Firsthand is not affiliated with its authors, and the site does not state that any model was trained on Firsthand data.

Can Firsthand collect egocentric data for a VLA program?

The site states custom egocentric collection to a written spec, fast and on demand, with 21-joint 3D hand pose, calibrated stereo depth, 200 Hz IMUs and one task instruction per episode, confirmed on a Spec Pilot.