Recoleta
GitHub
Source note

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

Authors
Shahram Najam Syed, Arthur Jakobsson, Haoran Hao, Jeffrey Ichnowski
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot Foundation ModelWorld ModelDynamic ManipulationSim2real
Open arXivMarkdown source

Summary

AHEAD is a wrapper for frozen VLA robot policies that predicts task-relevant future visual tokens before action decoding. The paper claims large gains on moving-object manipulation in simulation and on a UFactory xArm 7 without retraining the 7B OpenVLA base model.

Problem

  • Current VLA policies act on the latest camera frame, so their actions target stale object positions when objects move during execution.
  • This matters for conveyor grasping, rolling balls, projectile catching, handovers, and other tasks where reaction-only control loses the intercept window.
  • The target is real-time dynamic manipulation while keeping the pretrained VLA vision encoder, language encoder, and action decoder frozen.

Approach

  • AHEAD adds a 4.9M-parameter latent world model around frozen 7B OpenVLA.
  • RAFT optical flow estimates per-patch velocity and acceleration from recent frames.
  • A language-and-motion mask selects only task-relevant or moving patch tokens, usually 30 to 60 tokens out of 196.
  • A conditional flow-matching model predicts future VLA feature tokens in latent space, using constant-acceleration updates for motion conditioning.
  • The rollout stops when sample variance crosses an uncertainty threshold, with Kmax=10, S=5 samples, 5 Euler steps, and typical realized horizons of 3 to 5 steps.

Results

  • In 20 dynamic simulation scenarios, AHEAD reports 79% to 97% success; the strongest baseline reports 31% to 58%.
  • On constant-velocity and acceleration/deceleration simulation tasks, AHEAD reports 87.7% to 97.3% success, while DreamVLA reaches 30.7% to 58.3% depending on the task.
  • In a conveyor speed sweep from 0 to 40 cm/s, AHEAD stays at 95.4% to 97.6% success; DreamVLA drops from 96.8% at 0 cm/s to 47.2% at 40 cm/s.
  • On eight complex simulation tasks, AHEAD reports 79.4% to 95.8% success; the best baseline reaches at most 60.2%.
  • In the occlusion simulation task, AHEAD reports 79.4% success; DreamVLA, Open-loop VLA, Retargeting VLA, VLA + Fast Replan, and Streaming Diffusion Policy report 0.0%.
  • On physical xArm 7 tasks, AHEAD reports 29/30 to 30/30 on three conveyor or rolling-ball tasks, 23/30 on paddle interception, and 19/30 on projectile catching where every baseline scores 0/30.

Link