MiniWAM

Figure S1RolloutsLIBERO · LIBERO-Plus · RoboTwin 2.0

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.

(a) LIBERO standard evaluation

Pick up the black bowl between the plate and the ramekin and place it on the plate.LIBERO-Spatial
Open the top drawer and put the bowl inside.LIBERO-Goal
Put the yellow and white mug in the microwave and close it.LIBERO-10

(b) LIBERO-Plus perturbed scenes

Put both the alphabet soup and the tomato sauce in the basket.Camera viewpoints
Pick up the milk and place it in the basket.Light conditions
Turn on the stove and put the moka pot on it.Background textures

(c) RoboTwin 2.0 clean split

Place the burger and fries on the tray.
Open the laptop.
Place both cans in the plastic box.

(d) RoboTwin 2.0 randomized split

Stack the green block on the red block.
Place both shoes in the shoebox.
Put the bread in the skillet.

Figure S2 · Tables S1–S2NeighborsNative vs. PRISM on unseen LIBERO-90 tasks

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

Querytransition
DINOv3 nativenearest neighbor
DINOv3 PRISMnearest neighbor
Step 0 / 16

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.

Table S1. Top-1 neighbors for both backbones retrieved task and scene

Representation Retrieved task Scene
(a) Open the microwaveKitchen 7
DINOv3 nativePut the white bowl on the plateKitchen 7same scene
DINOv3 PRISMClose the top drawer of the cabinet and put the black bowl on top of itKitchen 10
Wan-VAE nativePut the white bowl on the plateKitchen 7same scene
Wan-VAE PRISMPut the white bowl on the plateKitchen 7same scene
(b) Put the right moka pot on the stoveKitchen 8
DINOv3 nativeTurn off the stoveKitchen 8same scene
DINOv3 PRISMTurn off the stoveKitchen 8same scene
Wan-VAE nativeTurn on the stove and put the frying pan on itKitchen 3
Wan-VAE PRISMTurn off the stoveKitchen 8same scene
(c) Pick up the tomato sauce and put it in the basketLiving Room 1
DINOv3 nativePick up the cream cheese box and put it in the basketLiving Room 1same scene
DINOv3 PRISMPick up the orange juice and put it in the basketLiving Room 2
Wan-VAE nativePick up the alphabet soup and put it in the basketLiving Room 1same scene
Wan-VAE PRISMPick up the orange juice and put it in the basketLiving Room 2
(d) Pick up the red mug and place it to the right of the caddyStudy 3
DINOv3 nativePick up the white mug and place it to the right of the caddyStudy 3same scene
DINOv3 PRISMPick up the book and place it in the right compartment of the caddyStudy 1
Wan-VAE nativePick up the white mug and place it to the right of the caddyStudy 3same scene
Wan-VAE PRISMPick up the yellow and white mug and place it to the right of the caddyStudy 1

Across the four queries, 7 of 8 native neighbors come from the query's own scene, against 3 of 8 PRISM neighbors.

Table S2. All 882 queries Table 6 of the paper, native → PRISM

Metric Main view Wrist view
Jaccard@5 ↑0.170 → 0.5120.107 → 0.333
Action RMSE ↓0.400 → 0.3650.370 → 0.272
EEF error (m) ↓0.083 → 0.0720.089 → 0.076
Gripper agr. ↑0.758 → 0.7930.820 → 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.

Retrieval protocol

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.

Figure S3PredictionsPredicted vs. recorded video

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.

Pick up the black bowl between the plate and the ramekin and place it on the plate.LIBERO-Spatial
Pick up the black bowl from table center and place it on the plate.LIBERO-Spatial
Pick up the alphabet soup and place it in the basket.LIBERO-Object
Pick up the butter and place it in the basket.LIBERO-Object
Open the middle drawer of the cabinet.LIBERO-Goal
Put the bowl on the plate.LIBERO-Goal
Turn on the stove and put the moka pot on it.LIBERO-10
Put the yellow and white mug in the microwave and close it.LIBERO-10

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 S4Real robotDual-arm Franka · head camera

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.

(a) Paired trials MiniWAM vs. native co-training on the same scene

(b) Lighting disturbances unseen during training

Place the light blue cup into the box.Room lighting changes during the episode
Place the green tape into the box.Room lighting changes during the episode

(c) Recovery retrying after a failed first attempt

Place the green tape into the box.Misses the first grasp, then retries
Place the light blue cup into the box.Misses the first grasp, then retries