Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control
Enfold uses a world model’s internal future-building states for control, then skips running the generator at deployment.
The paper trains a current-only encoder to match multi-level states from a generator as it processes observed futures. That representation conditions future generation during training and feeds task heads, while task gradients are blocked from reshaping the encoder. In tests across LIBERO, RoboTwin2.0, and real-robot tasks, Enfold cuts action latency by 3.7x versus Fast-WAM, while Enfold-Flash reaches 10.1x. The authors report that the system adapts when a human changes the current scene, arguing against simple fixed-trajectory replay. HF Daily Papers' note
The paper trains a current-only encoder to match multi-level states from a generator as it processes observed futures. That representation conditions future generation during training and feeds task heads, while task gradients are blocked from reshaping the encoder. In tests across LIBERO, RoboTwin2.0, and real-robot tasks, Enfold cuts action latency by 3.7x versus Fast-WAM, while Enfold-Flash reaches 10.1x. The authors report that the system adapts when a human changes the current scene, arguing against simple fixed-trajectory replay. HF Daily Papers' note
score 5