KAIST's FRED lifts robot policy success from 19.4% to 50% via self-distilled future video representations
kaist-ai · hf · 2026-10-23
KAIST and RLWRLD introduce FRED (Future REpresentation learning via self-Distillation), improving single-step future video representations in World Action Models, where action decoders rely directly on representations from one denoising step.
- Method: self-distillation with targets extracted from ground-truth future observations, plus temporal conditioning on historical observations.
- Results: consistently outperforms competing WAMs across real-world manipulation, with the largest gains on long-horizon and out-of-domain zero-shot tasks; human-to-robot skill transfer reaches 50.0% success on tasks absent from robot training data (shown only in human videos) versus 19.4% for the best competing WAM.
- Code and models are open-sourced on GitHub and Hugging Face.
More from Embodied
- 10 of 12 Beta Robots Worked Without Fine-Tuning — Do We Still Need It? — micoolcho · 2026-10-23
- EmbodiedEvo lets robot agents self-train, banking reusable Skills and Lessons across tasks — jiqizhixin · 2026-10-23
- Figure prints 81 minutes of robot data per second, spent ~$8.1M on data last week — joecole · 2026-10-23
- Just three physical AI agents are actually paid for at scale, argues investor — jokrvivek · 2026-10-23
- UCLA roommates-turned-founders: the story behind sidewalk delivery robot startup Coco Robotics — zhoubolei · 2026-10-23
- Figure 04 hits full design lock; Brett Adcock calls it the iPhone 1 of humanoids — chris_j_paxton · 2026-10-23