Rethinking Actions: Continuity-Aware Priors for World-Model Planning
Organizations: Beijing University of Posts and Telecommunications · Alibaba Group
Abstract
Actions in expert demonstrations exhibit temporal structure. An ordered-versus-shuffled analysis across five environments confirms this dependence. We ask whether action proposals that better reflect expert continuity can improve visual world-model planning. Rather than imposing generic smoothness, we learn continuations from the temporal structure of demonstrated actions. Our locality-gated deep-kernel Gaussian-process prior, LG-DKGP, combines recent and longer-range executed action history, predicts a neural continuation, and learns a temporal Gaussian-process (GP) residual. A training-derived exploration scale initializes search around the corrected mean; a frozen world model evaluates progress toward the goal. Training first fits the neural mean, then jointly optimizes the prior with the GP residual using expert future actions. The resulting planner achieves 87.07% macro success across five simulated tasks spanning native and fragment-level goal attainment, compared with 71.03% for INTACT, 61.66% for PRISM, and 58.66% for LeWM. Component interventions and expert-relative action-structure comparisons support demonstration-informed continuations and reveal task-dependent effects of visual conditioning, history length, and planning horizon.
Figures & tables
| Task | Episodes | 95% interval | ||
|---|---|---|---|---|
| PushT | 1842 | 0.354 | 0.815 | [0.805, 0.825] |
| Cube | 973 | 0.360 | 0.866 | [0.859, 0.873] |
| Caddy | 1000 | 0.267 | 0.855 | [0.850, 0.859] |
| Scene2 | 985 | 0.299 | 0.746 | [0.742, 0.750] |
| RoboTwin | 1000 | 0.020 | 0.514 | [0.502, 0.526] |
| Method | PushT | Cube | Caddy | Scene2 | RoboTwin | Macro |
|---|---|---|---|---|---|---|
| LeWM | 58.66 | |||||
| iCEM | 59.33 | |||||
| INTACT | 71.03 | |||||
| PRISM | 61.66 | |||||
| LG-DKGP (Ours) | 87.07 |
| Task | Fixed | Diag. | +Bounds | LG-DKGP |
|---|---|---|---|---|
| PushT | 97.60 | 97.60 | 97.60 | 97.60 |
| Cube | 99.20 | 98.80 | 98.80 | 98.80 |
| Caddy | 83.34 | 76.97 | 76.98 | 83.32 |
| Scene2 | 75.56 | 69.44 | 72.98 | 76.96 |
| RoboTwin | 18.88 | 72.86 | 77.81 | 78.66 |
| Macro | 74.92 | 83.14 | 84.84 | 87.07 |
| Task | Closer | (%) | 95% interval | |
|---|---|---|---|---|
| PushT | 209/250 | 83.6 | ||
| Cube | 225/250 | 90.0 | ||
| Caddy | 221/222 | 99.5 | ||
| Scene2 | 222/226 | 98.2 | ||
| RoboTwin | 219/219 | 100.0 |
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
| PushT | Cube | Caddy | Scene2 | RoboTwin | Macro | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Variant | Mean SD | Mean SD | Mean SD | Mean SD | Mean SD | Mean | ||||||
| LG-DKGP (Ours) | — | — | — | — | — | 87.07 | — | |||||
| History only | 78.21 | |||||||||||
| Linear | 55.58 | |||||||||||
| w/o GP residual | 83.27 | |||||||||||
| Fixed gate | 86.30 | |||||||||||
| Task | Mean only | CEM | |
|---|---|---|---|
| PushT | +39.20 | ||
| Cube | +0.40 | ||
| Caddy | +22.50 | ||
| Scene2 | +24.01 | ||
| RoboTwin | +3.51 |
| History | |||
|---|---|---|---|
| 0 | |||
| 1 | |||
| 5 | |||
| 16 | |||
| 32 |
| Goal offset | PushT | Cube | Caddy | Scene2 | RoboTwin |
|---|---|---|---|---|---|
| 25 | |||||
| 30 | |||||
| 35 | |||||
| 40 | |||||
| 45 | |||||
| 50 |
| Task | Method | MMD | Profile error | Successes / eligible | Success (%) |
|---|---|---|---|---|---|
| PushT | LG-DKGP | ||||
| PushT | LeWM | ||||
| Cube | LG-DKGP | ||||
| Cube | LeWM | ||||
| Caddy | LG-DKGP | ||||
| Caddy | LeWM |
| Goal offset | PushT | Cube | Caddy | Scene2 | RoboTwin |
|---|---|---|---|---|---|
| 30 | |||||
| 35 | |||||
| 40 | |||||
| 45 | |||||
| 50 |
| Task | Raw scalar | Cal. scalar | Cal. box | NLL |
|---|---|---|---|---|
| PushT | 3.28 | 88.83 | 38.58 | 0.81 |
| Cube | 4.47 | 93.03 | 8.75 | 0.75 |
| Caddy | 7.93 | 95.49 | 33.72 | 0.82 |
| Scene2 | 9.72 | 96.33 | 41.01 | 0.90 |
| RoboTwin | 35.58 | 87.66 | 54.57 | 1.02 |
| Task | Fixed gate | Local only | Global only |
|---|---|---|---|
| PushT | 0.788 [0.506, 0.960] | 0.687 [0.349, 0.939] | 0.729 [0.427, 0.917] |
| Cube | 0.354 [0.150, 0.692] | 0.354 [0.150, 0.692] | 0.354 [0.150, 0.692] |
| Caddy | 0.634 [0.535, 0.730] | 0.583 [0.485, 0.686] | 0.494 [0.410, 0.584] |
| Scene2 | 0.591 [0.502, 0.687] | 0.544 [0.453, 0.630] | 0.435 [0.364, 0.505] |
| RoboTwin | 0.625 [0.535, 0.712] | 0.586 [0.503, 0.667] | 0.554 [0.474, 0.637] |
| Task | Fixed | Expert diagonal | Diagonal + bounds | Expert isotropic |
|---|---|---|---|---|
| PushT | ||||
| Cube | ||||
| Caddy | ||||
| Scene2 | ||||
| RoboTwin |
| Half bandwidth | Nominal | Double bandwidth | ||||
|---|---|---|---|---|---|---|
| Task | LeWM | LG-DKGP | LeWM | LG-DKGP | LeWM | LG-DKGP |
| PushT | 0.2541 | 0.0889 | 0.1257 | 0.0411 | 0.0299 | 0.0091 |
| Cube | 0.2913 | 0.0817 | 0.2624 | 0.0488 | 0.0790 | 0.0171 |
| Caddy | 0.4473 | 0.4037 | 0.5864 | 0.2709 | 0.2719 | 0.0648 |
| Scene2 | 0.4368 | 0.4306 | 0.6279 | 0.4799 | 0.3793 | 0.1745 |
| RoboTwin | 0.6249 | 0.5509 | 0.8449 | 0.3271 | 0.9779 | 0.0684 |
| Task | LG-DKGP | Constant | Shuffled |
|---|---|---|---|
| PushT | 0.0411 | 0.1460 | 0.1420 |
| Cube | 0.0488 | 0.1385 | 0.2149 |
| Caddy | 0.2709 | 0.1382 | 0.1855 |
| Scene2 | 0.4799 | 0.1501 | 0.1540 |
| RoboTwin | 0.3271 | 0.3033 | 0.4137 |
| Task | Method | Successful / failed | MMD 2 on success | MMD 2 on failure |
|---|---|---|---|---|
| PushT | LeWM | 234/16 | 0.1203 | 0.2043 |
| PushT | LG-DKGP | 244/6 | 0.0394 | 0.1112 |
| Cube | LeWM | 159/91 | 0.2187 | 0.3388 |
| Cube | LG-DKGP | 247/3 | 0.0492 | 0.0145 |
| Caddy | LeWM | 122/100 | 0.5728 | 0.6031 |
| Caddy | LG-DKGP | 184/38 | 0.2562 | 0.3420 |
| Task | Prior train | Validation |
|---|---|---|
| PushT | 14,733 | 1,842 |
| Cube | 7,785 | 973 |
| Caddy | 8,000 | 1,000 |
| Scene2 | 7,869 | 985 |
| RoboTwin | 8,000 | 1,000 |
| Module | Layers / representation |
|---|---|
| Action-history token | Action projection , valid-difference projection , increment-validity projection , and age positional encoding of width 128. |
| Conditional attention query | ; linear projection . |
| Gated attention (two layers) | Four heads, 32 features/head; learned nonnegative locality rates; local/global fusion . Gate MLP with GELU and sigmoid; initial gate 0.25. |
| Layer update | Residual + LN; FF with GELU and dropout; residual + LN. All-invalid history uses a learned 128-dimensional fallback. |
| Neural mean | ; GELU; LN; . Shared across forecast leads. |
| Query feature | ; GELU; LN; . |
| Task | Views | Action block | ||
|---|---|---|---|---|
| PushT | 2 | 1 | 192 | 10 |
| Cube | 5 | 1 | 192 | 25 |
| Caddy | 7 | 2 | 512 | 35 |
| Scene2 | 7 | 2 | 512 | 35 |
| RoboTwin | 14 | 3 | 512 | 70 |
| Study | Settings and comparisons | Seeds |
|---|---|---|
| Main performance | , ; 250/250/222/226/219 eligible queries in task order. | 5 |
| Method/component study | Same task allocation; History only, Linear, fixed gate 0.25, local-only 0, global-only 1, residual removal, raw-kernel refit. | 5 |
| History/schedule | RoboTwin; ; ; ; 41/46/46 eligible queries per configuration. | 3 |
| Fixed goal offsets | All five tasks; offset ; , , ; offset-specific eligibility. | 3 |
| Mean execution / search | Posterior-mean commands versus candidate optimization; retained per-task history interfaces as detailed in Table S2 . | 5 |
| Search initialization | Fixed 0.3, expert diagonal, diagonal+bounds, expert isotropic; retained full-prior checkpoints, , . | 5 |
| Proposal standard deviation | PushT | Cube | Caddy | Scene2 | RoboTwin |
|---|---|---|---|---|---|
| Archived fixed-scale sensitivity | |||||
| 0 | |||||
| 0.1 | |||||
| 0.3 | |||||
| 0.5 | |||||
| 0.7 | |||||
World Model-Based Planning
World model-based planning uses models of how environments change to compare imagined actions and choose goal-directed behavior.
See all 302 papers