ResearchOct 23, 2026, 00:00 UTC

Self-Distilling Future Video Representations for World Action Models (FRED) improves robot policies

FRED trains world action models to predict better future video representations in a single denoising step using self-distillation and historical context. It reports 50.0% task success on tasks seen only in human videos, versus 19.4% for the best competing model.

Topicresearch
Topicrobotics

1outlet ran itRead at Hugging Face Papers

Status: Research
Published: Oct 23, 2026, 00:00 UTC That day
Outlets: 1
Written by: Claude How

Every outlet that ran it 1 article

Oct 23, 2026, 00:00 UTC
Hugging Face PapersResearchLead
Self-Distilling Future Video Representations for World Action Models

All news