Generated illustration for Learning by Watching Human Videos — Paper 31 of 41Seedream 4.0 · GMI Studio

Illustration brief the model wrote

A large monitor displays a human hands assembling a bookshelf from an instructional video. A robot arm below the monitor reaches towards a pile of boards. Minimalist lab setting, robot arm matching the hand pose shown on screen.

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 2e8becab…), converted to PNG locally. Summaries can be wrong; the linked paper is the source of truth.

Atlas/Skild AI·Physical AI company·2026

Learning by Watching Human Videos

Model tl;dr

Finetunes omni-bodied robot model on human videos and minimal robot data to acquire new skills, bypassing teleoperation's diversity and scale limits.

human videorobot learningomni-bodied modelimitation learningteleoperation bottleneckskill acquisition

What changed

Model

Before, robot skill acquisition relied on large amounts of teleoperated data, which was costly and limited in diversity and scale. This paper shows that finetuning a base model on human video with less than an hour of robot data can teach new skills, leveraging abundant internet human videos.

So what, in plain English

Model

Imagine a robot that learns to cook by watching YouTube cooking shows, not by someone moving its arms for hours. This lets robots learn many skills quickly from videos people already make.

Contribution to industry

Model

This paper contributes a method to reduce the need for expensive teleoperation data collection, enabling robots to learn from online human videos. It accelerates skill acquisition for consumer and industrial robots, and Skild AI's omni-bodied model can be used by robotics companies to quickly add new skills without extensive human demonstration.

Typed judgments

Jev

★★★☆☆ robots · Physical AI 95%

Real hardware 51% · Open code / data 8% · Foundation-model based 81% · Survey / benchmark-only 7%

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. Finetunes the omni-bodied model on human video — egocentric footage and instructional video — with under an hour of robot data to acquire new skills, tackling the diversity and scale limits of teleoperation.

Industry. Frames internet-scale human video as the robot data flywheel, potentially relieving the teleoperation bottleneck that caps how fast the whole industry can add skills.

Read the paper ↗Official technical research post ↗All 3 Skild AI papers →