Skip to content
RobotWorld
← Back to Videos

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

Loading video
Loading video

The same VLA was pre trained on 100 hours of the OTS dataset from Human Archive and on 100 hours randomly sampled from each of the microagi and MeckaAI datasets, then evaluated on EgoVerse. Human Archive data achieved the strongest results with the highest overall task success rate, and showed particularly strong generalization to unseen objects.

VLAPre-training DataAblationEgoVerseHuman ArchiveRobot LearningEmbodied AIDatasets
Category: research
Author: @babugi28
Date: 2026-10-15T00:00:00
Duration: 42.9s
Source: https://x.com/babugi28/status/2110787571550994777

Related Videos

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 · 372 views · 2026-10-08
World-action modelsVLAReal-time control
Reka Releases Rho-1, a 19B Omni Model
1:00

Reka Releases Rho-1, a 19B Omni Model

@RekaAILabs · 428 views · 2026-10-05
Embodied AIOmni ModelMultimodal
Gen-HumanEgo: Open-Source Egocentric Human Data for Robot Learning
4:14

Gen-HumanEgo: Open-Source Egocentric Human Data for Robot Learning

@GenrobotAI · 968 views · 2026-09-23
Egocentric VisionRobot LearningOpen Source
FLUX 3 Action: Open-Weights 7B World Action Model
0:05

FLUX 3 Action: Open-Weights 7B World Action Model

@HuggingPapers · 760 views · 2026-09-23
World-action modelsVLARobot Learning
MessyMem: Persistent Spatial Memory for Robots
0:48

MessyMem: Persistent Spatial Memory for Robots

@anuvabanwasi · 855 views · 2026-09-16
Persistent MemoryMobile Manipulation3D scene graphs
Odyssey-3: A Foundation World Model for Physical Agents
1:37

Odyssey-3: A Foundation World Model for Physical Agents

@odysseyml · 1.2k views · 2026-09-15
World ModelsFoundation ModelsPhysical AI

As an Amazon Associate, we earn from qualifying purchases.