Intercepting the Future

Abstract

Vision-Language-Action models generalize across static manipulation but fail when objects move during execution. They map the current observation to an action and assume the scene is stationary, so latency exceeds the time available to grasp.

AHEAD wraps a frozen VLA with a motion-aware latent world model that forecasts future patch tokens from optical-flow kinematics, concentrates compute on language-and-motion salient patches, and feeds the prediction to the unchanged action decoder.

Method

A 4.9M-parameter predict-then-act wrapper around a frozen 7B OpenVLA. RAFT flow gives per-patch velocity and acceleration; a language-and-motion mask keeps task-relevant tokens; conditional flow matching rolls them forward until uncertainty says stop. The frozen action decoder never sees the stale frame (~158 ms per step).

AHEAD pipeline: optical flow, saliency, world model, frozen VLA decoder.

Simulation results

Successes first (79-97% in-distribution), then lower-success cases where prediction degrades.

Success Conveyor Constant-velocity pick. AHEAD stays above 95% as speed rises; baselines collapse.

Success Moving beam Kinematic update tracks the beam; static-scene VLAs miss.

Success Pushed object Human displaces the cup mid-reach; AHEAD retargets to the new intercept.

Success Hit the puck Reactive contact at a predicted intercept (air-hockey style).

Success Catch the ball Ballistic intercept with a net before bounce.

Success Pick & place on conveyor Place into a moving target box.

Success Multi-object conveyor Select the right moving item among distractors.

Success Catch the named ball Language chooses which ball; motion still predicted.

Success Occluded catch Keep forecasting while the target is hidden (~79%).

Success Mid-flight deflection Recover after an in-air redirection (still in-distribution).

Success Language + multi-object sort Ground the instruction while objects move.

Lower success Multiple deflection Unpredictable redirections; only with a long reaction window.

Lower success Occlusion deflection Trajectory changes while occluded; nothing to track mid-hide.

Lower success Chaotic Newton's cradle Cascading collisions scramble the trajectory.

Lower success Plinko drop Pegs randomize the path after each collision.

Conveyor belt: AHEAD vs baseline

Same scene, same seed. Open-loop VLA grasps where the object was; AHEAD contacts where it will be.

Open-loop VLA: fail

AHEAD: success

Results

79-97%

sim success, 20 dynamic scenes
(best baseline 31-58%)

30/30

physical conveyor place
(best baseline 12/30)

23/30

paddle intercept
(all baselines 0/30)

19/30

projectile catch
(all baselines 0/30)

4.9M

params, frozen 7B OpenVLA
~158 ms / step

BibTeX

@inproceedings{syed2026ahead,
  title     = {Intercepting the Future: Latent-Space Predictive World Model for Dynamic {VLA} Manipulation},
  author    = {Syed, Shahram Najam and Jakobsson, Arthur and Hao, Haoran and Ichnowski, Jeffrey},
  booktitle = {Proceedings of the 10th Conference on Robot Learning ({CoRL} 2026)},
  address   = {Austin, TX, USA},
  year      = {2026},
  eprint    = {2606.02486},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url       = {https://arxiv.org/abs/2606.02486}
}

Intercepting the Future: Latent-Space Predictive World Model for Dynamic VLA Manipulation