Manipulating moving objects requires a policy that can anticipate contact events, yet vision-language-action (VLA) policies are usually fine‑tuned only on the current observation. World action models (WAMs) learn predictive dynamics, but running a video‑scale teacher or explicitly imagining future frames at deployment is computationally expensive.
ForeTime-VLA introduces a dense $\pi_{0.5}$ policy that remains causal at inference while distilling a future‑aware, action‑equivalent representation from a frozen Fast‑WAM teacher. Offline, current and future video latents are compressed into a whitened 64‑dimensional target vector. Online, an eight‑frame history encoder predicts this target together with the manipulation phase and a normalized time‑to‑transition. Four future tokens and one phase token condition the VLM prefix, and the predicted future vector plus transition horizon condition the action expert. Training retains the original flow‑matching action target and adds cosine similarity, relational geometry, phase, time‑to‑transition, and action‑equivalence losses.
On a deduplicated conveyor‑belt dataset, 40k‑step checkpoints are evaluated on 768 matched windows per split. Test MAE drops from 0.134119 to 0.130593 (2.63% improvement, 95% CI: 0.82‑4.48%), L2 error decreases by 3.02%, with an inference latency increase of 2.46‑2.93%. Real‑robot experiments show ForeTime-VLA achieves 81.1% success on stationary grasps and 58.9% on slow‑moving grasps, surpassing the next best baseline by 12.2 and 22.2 percentage points respectively. Across three belt speeds it completes 44/90 grasps versus 23/90 for $\pi_{0.5}$, including 11/30 versus 2/30 at high speed. The offline orientation gains align with reduced real‑robot contact‑pose failures, confirming that causal future‑token distillation effectively improves dynamic manipulation without deploying the world‑model teacher.
Blogger's Review: ForeTime-VLA leverages compact future tokens to provide forward‑looking predictions in dynamic scenes, achieving higher grasp success while keeping latency low, and demonstrates the practical value of causal distillation for robotic manipulation.