GE-Act 2.0 is a world action model from AgiBot, the Shanghai robotics company, and the 3-minute-49-second upload sets out how it was trained and how it was tested. The description says the model was trained “entirely from scratch on embodied manipulation data,” with “No inherited video generators.” AgiBot reports task success rising from 17.1% to 44.1% as training data scaled from 300 to 30,000 hours. The figures come from AgiBot alone.

Towels, paper cups and 100 atomic tasks

The description says visual representation, future generation and action prediction were all trained “from random initialization,” with “No task-specific fine-tuning.” It describes a “real-robot zero-shot test” covering “unseen scenes, unseen objects, 100 atomic tasks, 20 skill categories, and two robot embodiments.”

The narration lists what was tested. “Across 100 tracking tasks, we tested atomic skills like” pick, place, stack, pour, press, open and close, it says. The description names four skills it says emerge at scale: folding towels, nesting paper cups, uncapping pens and arranging flowers.

The stated numbers: data scaled from 300 to 30,000 hours; success on G1-OP from 17.1% to 44.1%; a second embodiment, G2-90D, trained on “under 2% of the training data,” gaining 17.7 percentage points. The description says 2,000 hours of the training set are failed manipulations and deployment rollouts, and the narration says that “Most models discard a failed grasp.”

The narration also describes the architecture: a control-oriented autoencoder, a visual planner that “imagines several futures,” an inverse dynamics model, and a planner that “denoises in one pass and the whole system trains end to end.”

From GO-1 to a from-scratch world model

The transcript sets this against AgiBot’s own earlier model: “Last year, GE act showed that visual futures can guide robot action, but most world action models still inherit pre-trained video generators.” GE-Act 2.0 is presented as the version that drops the inheritance.

The line behind it is short. GO-1, AgiBot’s Vision-Language-Latent-Action model, was introduced on 11 March 2025, and the company reported that “Testing across five tasks showed GO-1 increasing success rates by 32% (46% to 78%).” A Chinese-language press release from August 2026 says the G2 robots AgiBot entered at the 2026 World Humanoid Robot Games ran an operating system built on its ViLLA foundation model GO, a GE world model and a distributed RW-RL system. That is the closest confirmable prior: the GE line already runs on robots the company says are in deployment.

What 3 minutes 49 seconds leave out

The upload does not say how much of the footage is autonomous and how much, if any, is teleoperated. It does not state playback speed, and the video is edited, so no run can be assumed continuous. Nothing states how many attempts sit behind each successful clip, or which of the 100 tasks the footage represents. There is no battery figure, no cost, no release date, and no statement of whether GE-Act 2.0 will be published, licensed or kept internal. Every number here comes from AgiBot and has not been independently tested.

Stated and not stated

• Stated: training data scaled “from 300 to 30,000 hours,” a 100× increase.
• Stated: task success on G1-OP rises “from 17.1% to 44.1%” with “no sign of saturation.”
• Stated: a second embodiment, G2-90D, gains “17.7 percentage points” on under 2% of the training data.
• Stated: “2,000 hours of failed manipulations and deployment rollouts” are used as training material.
• Not stated: whether the demonstrations were autonomous or teleoperated, at what playback speed, or how many takes.
• Not stated: any paper, code release, licence, price or availability for GE-Act 2.0.

Whether the curve holds past 30,000 hours

Every large humanoid team is chasing one model that transfers across tasks and bodies. AgiBot’s listed rivals include “Unitree Robotics, UBTech Robotics, Galbot, Fourier Intelligence,” along with “Tesla (Optimus), Figure AI, 1X Technologies.” What separates the claims is whether a curve like 17.1% to 44.1% keeps climbing, and whether it holds on hardware outside the lab that produced it.

For this to carry weight, the scaling trend would have to continue past 30,000 hours, the cross-embodiment gain would have to reproduce on robots the company did not build, and an outside group would have to run the same 100-task benchmark. A published method, a benchmark others can rerun, or deployment data from the G2 fleet would each say more than another edited demo.

Source: AGIBOT GE-Act 2.0 — the first native World Action Model to validate a pretraining and scaling path by AgiBot , published 2026-09-11. This article was written automatically from the video above and has not been independently verified.