OS3

Version v1.0
Footage
Company OS3 Inc. Blog deep dives
OS3
Versionv1.0
Footage2026-08-10
CompanyOS3 Inc.
Blogdeep dives

MODEL //

ONE PLANNER. ONE EXECUTOR.

Two layers with different jobs: a deliberate system that comprehends multi-step work, and a video-pretrained system that turns each step into motion.

HUMAN INSTRUCTION + SCENE SYSTEM 01 · DELIBERATE PLANNER grounds the instruction in the scene decomposes it into bounded steps visually verifies progress bounded step + goal SYSTEM 02 · EXECUTOR video-pretrained action model images + robot state + step target → time-indexed action chunk validated targets CONTROL STACK gates · chunk queue · motion shaping · actuator bridges H.A.L.E. 1.0 measured state · failures RL ON OUR HARDWARE targets failures pretraining: 150,000 hours of video · 2,000+ real-world tasks human first-person demonstrations before any robot data
fig. 01 · the two-system architecture

The data

01 Pretraining: 150,000 hours of video spanning 2,000+ real-world tasks.

02 Grounding: first-person human demonstrations before any robot data.

03 RL on our own hardware, aimed at the failure modes deployments expose.

Where it works, domain by domain: use cases →

Technical report on training is coming soon.