A LITTLE IMAGINATION. A USEFUL FUTURE.
Adaptive Spatiotemporal Imagination
for Efficient Robot Control
One refinement field decides WHEN to stop, HOW FAR to denoise, and WHERE to refresh.
Elastic-WAM performs four-stage tea preparation.
Close-up of cup and teapot handling.
Close-up of cup and teapot handling.
World-Action Models improve robot control by predicting future visual dynamics, but iterative video denoising incurs substantial inference overhead. Recent approaches reduce this cost through partial future denoising, yet typically rely on fixed refinement budgets. Our empirical observations reveal that preferred refinement depths vary across tasks and states, while different future regions exhibit varying relevance to action generation. Motivated by these findings, we propose Elastic-WAM, which adaptively allocates world refinement computation across temporal denoising time and spatial future space. A unified spatiotemporal refinement demand field governs when to stop world refinement, how far to advance each denoising step, and which regions to refresh. We parameterize this field with a shared temporal basis and spatial amplitudes, and learn it through a three-stage action-grounded supervised procedure. At inference, the learned field guides recurrent sparse world refinement to improve task success while reducing inference overhead. Evaluations across diverse simulation benchmarks and real-world long-horizon manipulation demonstrate state-of-the-art control performance with reduced inference latency.
01 / METHOD
WHEN / STOPPING
Stop when cumulative World Commit probability reaches its threshold.
HOW FAR / PROGRESSION
Set denoising distance using the shared temporal profile.
WHERE / SELECTION
Refresh a fixed 96 of 196 future tokens per sparse update.
Three-stage controller training; pretrained WAM frozen.
02 / MODEL COMPARISONS
03 / QUANTITATIVE EVALUATION
| Method | Nominal | Lighting OOD | Object OOD | |||
|---|---|---|---|---|---|---|
| SR ↑ | PS ↑ | SR ↑ | PS ↑ | SR ↑ | PS ↑ | |
| Fast-WAM | 0.0 | 12.5 | 0.0 | 5.0 | 0.0 | 10.0 |
| Joint-WAM | 30.0 | 37.5 | 30.0 | 35.0 | 30.0 | 35.0 |
| Elastic-WAM (Ours) | 50.0 | 57.5 | 50.0 | 52.5 | 30.0 | 47.5 |
Table 7. On long-horizon tea preparation, Elastic-WAM achieves the highest full-task success rate (SR, %) under nominal and lighting OOD conditions, and the highest average stage completion (PS, %) across all settings. Under object OOD, it matches Joint-WAM’s SR while achieving higher PS.
SPATIAL SELECTION · ROBOCASA-GR1
| Refinement strategy | Success ↑ |
|---|---|
| Random WHERE, K = 96 | 53.83 |
| All-active, K = 128 | 55.17 |
| FirstKV | 54.50 |
| Learned WHERE, K = 96 | 57.42 |
Table 5 · Adaptive temporal control except FirstKV. K: tokens refreshed per sparse update after dense initialization.
TEMPORAL CONTROL · LIBERO-PLUS
| Controller, learned WHERE K = 96 | Success ↑ |
|---|---|
| WHERE-only · fixed progression & commit | 71.71 |
| Fixed step · adaptive commit | 72.31 |
| Adaptive progression & commit | 73.71 |
+2.00 points over spatial selection alone.
Table 6(a) · Only temporal control changes.
TOKEN BUDGET · LIBERO-PLUS
K = 96 performs best among the tested budgets.
04 / QUALITATIVE ANALYSIS