This paper introduces Action--Future Coupled Effect Learning (AFCE), which shifts the prediction target of vision‑language‑action (VLA) policies from immediate action chunks to the physical Effect that will occur after execution. An offline tokenizer constructs temporally ordered Effect tokens from demonstrated actions, realized robot responses, and sparse visual transitions. Reconstruction objectives force the representation to retain executable detail and observed physical change.
A flow‑matching VLA predicts Effects from the current image, instruction, and robot state, while a state‑conditioned decoder translates them into actionable chunks. Embodiment‑specific interfaces enable both single‑arm and dual‑arm tasks to share the same Effect model.
On the DexJoCo benchmark, AFCE achieves an overall success rate of $50.97\%$, improving over $\pi_{0.5}$ and UVT by 11.82 and 15.21 percentage points respectively. It leads on all five bimanual tasks, highlighting the benefit of Effect supervision for coordinated dexterous manipulation. Project page: https://afceproject.github.io/AFCE-Web/.
Review