Self-Distilling Future Video Representations for World Action Models

One-line takeaway

KAIST's FRED targets World Action Models, robot policies that decode actions from future video representations taken from a single denoising step of a video diffusion model; it improves those representations via self-distillation with temporal context, since action quality depends directly on them.

KAIST 的 FRED 针对“世界动作模型”：这类机器人策略从视频扩散模型单步去噪得到的未来视频表示里解码动作，动作质量直接取决于这些表示；FRED 用带时间上下文的自蒸馏来学习更好的未来表示。