pico-WAM: Evaluating Counterfactual Prediction and Planning Error in a Compact World Action Model
Organizations: Khoury College of Computer Sciences Northeastern University
Abstract
When a World Action Model fails at control, is the bottleneck proposal generation or trajectory scoring? We introduce pico-WAM, a compact model with supplied physical state for tracing policy-dynamics interactions. Using 206 Push-T demonstrations and 15 training seeds, we compare data regimes and search configurations. Counterfactual branch training improves short-horizon cost ranking and enlarges the coverage advantage of policy-restarted over warm-started search by 4.8 percentage points; restart wins under both data regimes. Replacing learned trajectory scores with simulator costs improves coverage by 12.2 points while retaining learned proposals. Substituting simulator states in proposal generation alone shows no detected gain. Component diagnostics and a separate door-opening study show why better ranking need not yield better control or outperform imitation. These controlled interventions identify scoring as an effective improvement target in the tested setting and provide a reproducible protocol for diagnosing compact world models.
Figures & tables
| Training | Control | Coverage, % | Success, % |
|---|---|---|---|
| BC | Reactive | 78.3 3.3 | 23.3 |
| Current, demos | Reactive | 72.9 4.2 | 22.2 |
| Current, + branches | Reactive | 71.1 3.4 | 21.6 |
| WAM, demos | Reactive | 72.1 2.6 | 22.6 |
| Warm search | 21.3 2.9 | 0.0 | |
| Restart search | 43.8 4.1 | 1.8 |
| Measure | Demos | + branches | Difference |
|---|---|---|---|
| Prediction, 21 unseen test demonstrations | |||
| State nMSE, recorded, 15 steps | 0.011 | 0.012 | 0.000 [ 0.001, 0.002] |
| Block nMSE, random, 10 steps a | 0.025 | 0.007 | 0.019 [ 0.023, 0.015] |
| Block RMSE, px, random, 10 steps | 14.8 | 7.1 | 7.8 [ 9.3, 6.2] |
| Prediction, 50 fresh scripted-expert episodes | |||
| State nMSE, recorded, 15 steps | 0.042 | 0.020 | 0.022 [ 0.030, 0.014] |
| Variant minus full | Reactive | Restart |
|---|---|---|
| Auxiliary gradients blocked a | 5.1 [ 5.1, 15.3] | 44.3 [ 50.2, 38.4] |
| Image only | 31.5 [ 37.9, 25.1] | 41.6 [ 46.7, 36.4] |
| State only | 0.7 [ 8.0, 9.4] | 3.3 [ 7.1, 0.5] |
| No instruction | 1.5 [ 6.3, 9.3] | 4.1 [ 8.6, 0.5] |
| No future-state loss | 0.4 [ 5.2, 5.9] | 6.4 [ 16.6, 3.8] |
| No latent loss | 0.5 [ 8.2, 9.2] | 5.2 [ 14.3, 3.9] |
| Study, seeds | Reference | Coverage | Success |
|---|---|---|---|
| Matched, 15 | BC | 0.1 [ 3.1, 3.4] | 9.5 [3.4, 15.6] |
| Horizon-matched, 15 | BC | 0.2 [ 4.0, 4.4] | 10.0 [4.7, 15.3] |
| Original gym-pusht, 15 | BC | 1.4 [ 4.6, 1.7] | 8.1 [1.7, 14.5] |
| Tuned BC, 15 | tuned BC | 2.7 [ 0.6, 6.0] | 0.1 [ 7.3, 7.1] |
| Diffusion Policy, 5 | DP | 1.0 [ 7.6, 5.5] | 16.7 [ 27.9, 5.5] |
| State dynamics, 15 | latent | 1.9 [ 4.7, 1.0] | 12.0 [5.7, 18.3] |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
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 294 papers