Skip to content
RobotWorld
ZH
← Back to Videos

RT-2: Turning Vision and Language Directly Into Robot Actions

Loading video
Loading video

RT-2 represents robot actions as tokens so a vision-language model becomes the controller itself. Web-scale pretraining transfers into better generalization on unfamiliar objects and tasks, including simple semantic reasoning.

VLAVision-language-action (VLA)RT-2Robot LearningGoogle DeepMind
Category: research
Author: @Alacritic_Super
Date: 2026-09-22T00:00:00
Duration: 112.228s
Source: https://x.com/Alacritic_Super/status/2102395650109509796
Reference: https://arxiv.org/abs/2307.15818

Related Videos

A-0 Trains Robot Policies on 16,000 Hours of Human Video
0:39

A-0 Trains Robot Policies on 16,000 Hours of Human Video

@yh_kris · 68 views · 2026-11-04
Human demonstrationsRobot Learning
RL Post-Training Generalizes Robot Policies Beyond Seen Tasks
0:13

RL Post-Training Generalizes Robot Policies Beyond Seen Tasks

@MichalAKorniak · 60 views · 2026-11-05
Robot LearningRL Post-Training
Astra Edits Four Weights to Lift VLA Success
0:32

Astra Edits Four Weights to Lift VLA Success

@Lingxiao234 · 112 views · 2026-11-04
VLAWeight Editing
Genie 4: A Steerable World Model for Robot Training
0:15

Genie 4: A Steerable World Model for Robot Training

@TheHumanoidHub · 170 views · 2026-10-27
Genie 4World Models
Same VLA, three data sources: Human Archive OTS wins on task success
0:43

Same VLA, three data sources: Human Archive OTS wins on task success

@babugi28 · 203 views · 2026-10-15
VLAPre-training Data
Long-WAM: Scaling World-Action Model Context to 19.2s in Real Time
2:38

Long-WAM: Scaling World-Action Model Context to 19.2s in Real Time

@AaronWeiHuang · 549 views · 2026-10-08
World-action modelsVLA

As an Amazon Associate, we earn from qualifying purchases.