Skip to content
RobotWorld
← Back to Videos

T2-z Adapts World Models at Test Time

Loading video
Loading video

T2-z optimizes a 32D environment latent through world prediction at test time, lowering object error from 15.20 to 7.00 and raising action success from 33.73% to 72.96% across 7 simulations and 4 real tasks.

World ModelsTest-Time OptimizationEnvironment LatentsRobot LearningT2-zStanford
Category: research
Author: @yining_hong
Date: 2026-10-16T00:00:00
Duration: 160.681s
Source: https://x.com/yining_hong/status/2111244176021164472
Reference: https://arxiv.org/abs/2610.22364

Related Videos

DreamFAR: A Robot-Only World Model Trained from Scratch
1:14

DreamFAR: A Robot-Only World Model Trained from Scratch

@_YimingDou · 126 views · 2026-10-29
DreamFARWorld Models
Sam Altman on AI Scale, Utilities, and One-Person Labs
26:31

Sam Altman on AI Scale, Utilities, and One-Person Labs

@maarcoofdezz · 764 views · 2026-09-24
OpenAISam Altman
L3G-World Stores Scene Memory in 3D Gaussians
0:15

L3G-World Stores Scene Memory in 3D Gaussians

@Machinedelusion · 78 views · 2026-10-31
World Models3D Gaussian Splatting
Stanford After42 finds research questions by trying to solve them
0:25

Stanford After42 finds research questions by trying to solve them

@aigclink · 88 views · 2026-10-29
AI for ScienceResearch Agents
Genie 4: A Steerable World Model for Robot Training
0:15

Genie 4: A Steerable World Model for Robot Training

@TheHumanoidHub · 141 views · 2026-10-27
Genie 4World Models
JST-2: From a Prompt to an Interactive 3D World
1:00

JST-2: From a Prompt to an Interactive 3D World

@yyyole · 186 views · 2026-10-25
World Models3D Gaussian Splatting

As an Amazon Associate, we earn from qualifying purchases.