Vision-Language-Action or World-Action? Just Action Models
The VLA versus world-model debate bundles three separate choices. Separating them yields one recipe: a foresight loss whose head is discarded, cutting FLOPs by 73%.
The claim
A foresight loss lifts robot control success without keeping any generator at inference time.
The problem
The vision-language-action and world-action camps disagree about which backbone to use.
Their debate bundles three separable design decisions into one question.
Keeping future-loss machinery inside the policy costs compute at inference.
How it works
JAM pairs a pretrained visual backbone with an action expert.
It trains with an action loss plus a foresight loss.
The foresight head predicts frozen encodings of the recorded future.
That head is discarded after training ends.
No visual generation runs during training or inference.
Figures
What changes
JAM beats matched FastWAM and ImageWAM policies on LIBERO and LIBERO-Plus.
LIBERO-Plus success rises from 55.3% to 64.5%.
Cached foresight targets cut estimated training FLOPs per update by 73%.
Limits
Gains are measured on three benchmark suites.
RoboTwin success improves only slightly over the baseline.
Real robot transfer is not reported in the abstract.
Source · arXiv