---
type: "post"
id: "agibot-unveils-go-3-preview-embodied-foundation-model-e20dc4"
slug: "agibot-unveils-go-3-preview-embodied-foundation-model-e20dc4"
title: "AgiBot unveils GO-3 Preview embodied foundation model built on WRAM, plans full open-source release by end of 2026"
description: "AgiBot used its 2026 Innovation Day to introduce GO-3 Preview, an embodied foundation model built around joint task reasoning, future-state prediction, and robot action generation."
retrieval_nugget: "AgiBot used its 2026 Innovation Day to introduce GO-3 Preview, an embodied foundation model built around joint task reasoning, future-state prediction, and robot action generation."
published_at: "2026-10-30"
updated_at: "2026-10-30"
record_date: "2026-10-30"
date_kind: "published_at"
topics: ["google","gemini"]
entities: ["AgiBot","GO-3 Preview","Link-U OS 1.0","WRAM","AHA3","AGILE","OpenAtom Foundation","Worken"]
related_patterns: ["harness-architecture-outlives-model-choice","agent-economics-moves-to-completed-work"]
novelty_label: "same_event"
what_changed: "Same event as AGIBOT unveils GO-3 Preview embodied foundation model at Innovation Day 2026, plans open-source release by end of 2026"
prior_refs: ["memory:610d41c3-6627-8036-a5db-02345fc4665c","memory:5782dcb6-db00-8540-a572-02c318c13b8b"]
source_url: "https://www.agibot.com/article/231/detail/131.html"
source_urls: ["https://www.agibot.com/article/231/detail/131.html"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:agibot-unveils-go-3-preview-embodied-foundation-model-e20dc4"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/agibot-unveils-go-3-preview-embodied-foundation-model-e20dc4/","markdown":"https://newruntime.com/posts/agibot-unveils-go-3-preview-embodied-foundation-model-e20dc4.md","json":"https://newruntime.com/posts/agibot-unveils-go-3-preview-embodied-foundation-model-e20dc4.json"}
---

# AgiBot unveils GO-3 Preview embodied foundation model built on WRAM, plans full open-source release by end of 2026

## Retrieval answer

AgiBot used its 2026 Innovation Day to introduce GO-3 Preview, an embodied foundation model built around joint task reasoning, future-state prediction, and robot action generation.

## Summary
AgiBot used its 2026 Innovation Day to introduce GO-3 Preview, an embodied foundation model built around joint task reasoning, future-state prediction, and robot action generation. It also launched Link-U OS 1.0 and says both the model and operating system are intended to become fully open source by the end of 2026, potentially creating a more inspectable starting point for robotics builders.

## What happened
AgiBot announced a broad set of robotics products and research efforts at its Innovation Day 2026 in Shanghai on October 29. The central software announcement was GO-3 Preview, which the company describes as its next-generation embodied AI foundation model. It is built on what AgiBot calls WRAM, short for World Reasoning Action Model. In the company's description, WRAM jointly handles task reasoning, prediction of future states, and generation of robot actions rather than treating those functions as separate systems.

The company says GO-3 was trained in several stages: pre-training; post-training using its Genie Evolver system, which combines real robots, simulation, and world models; and post-deployment learning. AgiBot reported that GO-3 ranked first in task-success rate and process score on RoboDojo, had the highest overall score on RoboTwin, and ranked first on RoboColiseum. Those are company-reported benchmark results; the announcement does not provide enough methodological detail to independently assess comparability, test conditions, or margins over other systems.

The access promise is as important as the preview itself. AgiBot says GO-3 Preview will be fully open sourced by the end of 2026. The announcement does not state a release date, licence, model size, supported robot embodiments, hardware requirements, or whether the release will include training data, training code, evaluation harnesses, and deployment components. For now, the item is a declared release plan, not an available package that a team can download and test.

AgiBot also formally launched Link-U OS 1.0, a robot operating system that it says is planned to become fully open source by year-end and be donated to the OpenAtom Foundation. The source presents this alongside AHA3, AgiBot's third-generation humanoid hardware architecture, and AGILE, its motion-intelligence framework. AHA3 combines foundations for locomotion, interaction, and task execution in one architecture, while AGILE spans motion foundation models, generative motion control, and whole-body perception and control. AgiBot also highlighted a new self-evolving agent capability intended to improve a robot's performance from experience in real environments.

The event framed these technical releases around deployment economics. AgiBot introduced Worken, or Effective Intelligent Work Hours, as a proposed productivity measure: active robots multiplied by autonomous working hours and an output-efficiency ratio. The ratio is meant to compare acceptable output on a defined task with a reference benchmark, such as human output under established operating standards. AgiBot calls Worken an evolving framework, to be refined through operational data rather than a finalized industry metric.

## What the archive adds
This is a new company and product event, but it belongs beside [Gemini Robotics Embodied Reasoning 2 entering public preview on Google Cloud](https://newruntime.com/posts/gemini-robotics-embodied-reasoning-2-enters-public-preview-e0fdff/). That earlier note covered enterprise availability for one member of Google's robotics model family; AgiBot's announcement instead centers on a future open-source model and robot operating system. It is a new fact in the broader embodied-model story, not a restatement.

## Trend signal
This supports [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/), cautiously. AgiBot is presenting a stack rather than a model in isolation: model training and post-deployment learning, motion intelligence, hardware architecture, operating system, and productivity measurement. The source does not show that these layers are interoperable or usable by outside teams yet, but its product framing recognizes that deployed robotics depends on more than model weights.

## Why now
AgiBot positions the announcement as robotics moving from manufacturing scale and isolated capability demonstrations toward sustained work in factories, warehouses, and commercial settings. Worken gives that framing a proposed metric: not robots built or powered on, but autonomous hours and acceptable task output. The promised open-source releases are presented within the same deployment-first strategy.

## Connections to other events
Google's Gemini Robotics family has also been split into specialized embodied capabilities, with Embodied Reasoning 2 later entering public preview on Google Cloud. AgiBot's approach differs in the source material: GO-3 combines reasoning, future-state prediction, and actions in one stated architecture, while Link-U OS is intended to supply an open operating-system layer for robots.

## Why it matters
For teams building physical AI, the practical question is not whether another company has shown a capable robot model. It is whether the model, runtime, hardware interfaces, evaluation method, and update loop can be inspected and adapted for a specific deployment. AgiBot is promising pieces of that stack, including an open model and OS, rather than only a hosted demonstration.

That could matter if the releases arrive with meaningful artefacts: weights, a clear licence, supported hardware targets, reproducible evaluation, and a usable path from simulation to deployed robots. In that case, builders could test whether a general embodied model reduces the need to train narrowly tailored policies for every task. Until then, the announcement is a direction of travel, not evidence of accessible developer infrastructure.

## What happens next
- **AgiBot releases source artefacts for GO-3 or clarifies its licence and supported hardware.** (3 months, 55%). Watch for: a repository, model card, weights, or explicit release terms.
- **Link-U OS receives an open-source release or OpenAtom Foundation handoff.** (3 months, 60%). Watch for: a public codebase, governance announcement, and installation documentation.
- **AgiBot publishes deployment data using Worken or task-specific benchmarks.** (6 months, 45%). Watch for: named sites, autonomous-hours figures, and defined output baselines.

## Limits and uncertainty
The public source is a company announcement, not an independent evaluation or a technical paper. Its benchmark leadership claims are vendor-reported, and it does not provide model size, dataset composition, safety performance, robot coverage, inference cost, latency, licence terms, or source-code availability. “Fully open sourced” is also not defined: the statement does not establish whether it means weights alone, training and evaluation code, data, or all of these. The company describes Worken as a proposal still to be refined, so it should not yet be treated as a comparable industry productivity standard.

## Skill to build now
Create an evidence checklist for any embodied-model evaluation your team considers: task definition, acceptable-output criteria, intervention count, autonomous operating time, hardware configuration, failure modes, and recovery procedure. AgiBot's Worken proposal is useful as a prompt to separate fleet size and powered-on time from verified useful work. That checklist will make it easier to judge later claims from GO-3, Link-U, or other robotics stacks.

## Reader takeaway
AgiBot is not merely previewing a robot model; it is proposing an open-stack path that joins model reasoning and action with an operating system and a measure of productive robot work. The meaningful test comes later: whether the promised releases include enough code, weights, licences, benchmarks, and deployment evidence for outside builders to reproduce useful work.

## Open question
Will AgiBot's eventual open-source release expose a genuinely reusable deployment stack, or only model artefacts that remain tightly coupled to its own robots and infrastructure?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)