Just Action Models

VLA or WAM? Three choices, unbundled.

A VLA ignores the recorded future, a WAM generates it, and JAM predicts its features.

The best target lies ahead.

Future-only offsets beat the current only.

Generation helps. Foresight helps more.

Six backbones, three families, three benchmarks: wherever generation is reported, foresight scores higher.

One fails, one finishes.

Six pairs play in sync from the same initial state: same backbone, different training.

Current features hint at what comes next.

Probes reconstruct the frame 16 steps ahead; with future targets, JAM (Ovis-U1) edges out ImageWAM.

JAM attends more to the second mug.

Paper example on Ovis-U1: the instruction names two mugs, and ImageWAM concentrates on the held one.

Against VLAs and WAMs.

JAM next to other policies on LIBERO-Plus, LIBERO and RoboTwin 2.0.

Nothing to generate, less to compute.

Generation denoises future frames every update; cached foresight processes only the current frame.

Start from a pretrained backbone, keep the target frozen.

A randomly initialized target still works, and robustness gains span six of seven perturbation types.

Cite Just Action Models.