Generated illustration for Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer — Paper 34 of 41Seedream 4.0 · GMI Studio

Illustration brief the model wrote

A humanoid robot standing in a modern living room. From above, dozens of first-person video cameras stream a montage of different people walking through various home interiors. Dashed lines connect those human paths directly to the robot's legs, showing imitation without robot demos. No text, no faces.

Provenance. Brief written by deepseek-ai/DeepSeek-V4-Flash via GMI Cloud on 2026-09-18 from the paper's title, primary link and the editor's notes; reviewed editorially. Graphic rendered by seedream-4-0-250828 through the GMI Studio request queue (request 588f1b17…), converted to PNG locally. Summaries can be wrong; the linked paper is the source of truth.

Atlas/Figure AI·Physical AI company·2025

Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer

Model tl;dr

Collects 100K+ home human egocentric videos; zero-shot transfers human navigation to humanoid robot without demos.

humanoidpretrainingegocentric videozero-shot transfernavigationdataset

What changed

Model

Before, humanoid robots required robot-specific demonstrations for learning navigation. This paper shows that internet-scale human egocentric video, collected via residential partnerships, enables zero-shot transfer of language-driven navigation to a humanoid, completely bypassing robot data collection.

So what, in plain English

Model

Imagine teaching a robot to navigate your house by watching thousands of people walk through their own homes on camera. This project makes that robot instantly able to move through a new home it has never seen, just by being told where to go in plain English.

Contribution to industry

Model

Figure AI's partnership with Brookfield provides an unprecedented real-home dataset covering over 100,000 residential units, establishing a new data strategy for humanoid pretraining. This zero-shot transfer capability could dramatically reduce the cost and time of deploying humanoid robots in real-world settings like homes and commercial facilities, enabling faster iteration from simulation to physical deployment.

Typed judgments

Jev

★★★☆☆ robots (also embodied) · Physical AI 91%

Real hardware 56% · Open code / data 6% · Foundation-model based 70% · Survey / benchmark-only 5%

Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.

Editor's notes

Summary. Builds an internet-scale humanoid pretraining dataset through a Brookfield partnership covering over 100,000 residential units, and shows zero-shot transfer from 100% egocentric human video to language-driven robot navigation with no robot demonstrations.

Industry. Establishes large human-video data partnerships across homes and commercial real estate as the strategy for humanoid pretraining.

Read the paper ↗Official technical report ↗All 4 Figure AI papers →