Learning Compact Future Targets for Efficient World-Action Modeling
Anonymous authors · Videos and tables for the author response
Figure S1. Successful closed-loop episodes of the 0.25B MiniWAM policy. The LIBERO-Plus perturbations and the RoboTwin 2.0 randomized scenes are unseen during training.
Figure S2. Query transitions from LIBERO-90 tasks unseen by the PRISM encoders, with their nearest neighbor from a different task under native features and under PRISM.
Main-view clips · top-1 neighbor per representation
Recorded LIBERO-90 demonstrations, frames t to t+16 at 6 frames/s. Neighbors come from different tasks in an episode-disjoint gallery; scenes may repeat.
| Representation | Retrieved task | Scene |
|---|---|---|
| (a) Open the microwaveKitchen 7 | ||
| DINOv3 native | Put the white bowl on the plate | Kitchen 7same scene |
| DINOv3 PRISM | Close the top drawer of the cabinet and put the black bowl on top of it | Kitchen 10 |
| Wan-VAE native | Put the white bowl on the plate | Kitchen 7same scene |
| Wan-VAE PRISM | Put the white bowl on the plate | Kitchen 7same scene |
| (b) Put the right moka pot on the stoveKitchen 8 | ||
| DINOv3 native | Turn off the stove | Kitchen 8same scene |
| DINOv3 PRISM | Turn off the stove | Kitchen 8same scene |
| Wan-VAE native | Turn on the stove and put the frying pan on it | Kitchen 3 |
| Wan-VAE PRISM | Turn off the stove | Kitchen 8same scene |
| (c) Pick up the tomato sauce and put it in the basketLiving Room 1 | ||
| DINOv3 native | Pick up the cream cheese box and put it in the basket | Living Room 1same scene |
| DINOv3 PRISM | Pick up the orange juice and put it in the basket | Living Room 2 |
| Wan-VAE native | Pick up the alphabet soup and put it in the basket | Living Room 1same scene |
| Wan-VAE PRISM | Pick up the orange juice and put it in the basket | Living Room 2 |
| (d) Pick up the red mug and place it to the right of the caddyStudy 3 | ||
| DINOv3 native | Pick up the white mug and place it to the right of the caddy | Study 3same scene |
| DINOv3 PRISM | Pick up the book and place it in the right compartment of the caddy | Study 1 |
| Wan-VAE native | Pick up the white mug and place it to the right of the caddy | Study 3same scene |
| Wan-VAE PRISM | Pick up the yellow and white mug and place it to the right of the caddy | Study 1 |
Across the four queries, 7 of 8 native neighbors come from the query's own scene, against 3 of 8 PRISM neighbors.
| Metric | Main view | Wrist view |
|---|---|---|
| Jaccard@5 ↑ | 0.170 → 0.512 | 0.107 → 0.333 |
| Action RMSE ↓ | 0.400 → 0.365 | 0.370 → 0.272 |
| EEF error (m) ↓ | 0.083 → 0.072 | 0.089 → 0.076 |
| Gripper agr. ↑ | 0.758 → 0.793 | 0.820 → 0.957 |
Leave-task-out retrieval over 882 episode-disjoint queries and 3,618 gallery transitions, top-5 neighbors. Jaccard@5 measures how often Wan-VAE and DINOv3 retrieve the same neighbors; the behavioral metrics compare each query with its neighbors, averaged over both backbones.
Neighbors are selected by cosine distance among 3,618 gallery transitions, excluding the query's task. The native readout spatially averages each frame's features; PRISM uses its full-grid posterior mean (3 × 4 × 4 tokens, 64 dimensions). Both see the observed future, so this tests representation geometry, not policy prediction. The four examples are illustrations.
The predictions are visually faithful, yet this policy is clearly weaker than PRISM on LIBERO-Plus-2K (48.5% vs. 58.1%, averaged over two seeds). Native co-training is therefore not limited by failing to predict the future.
Figure S3. Futures predicted by the native Wan-VAE world branch of the 0.25B model (native co-training, Table 1), decoded to RGB, against the recorded video.
Each clip shows one 16-step forecast, slowed from 20 to 5 frames per second and paused briefly on its last frame. Recorded frames are decoded from their VAE latents.
Figure S4. Real-robot rollouts of a dual-arm pick-and-place task: the left arm moves the black box to the center, and the right arm places the light blue cup or the green tape into it. MiniWAM and native co-training share the same 0.25B architecture and the same demonstrations, and run alternately on the same scenes. Head camera, played at 4× speed.