pico-WAM: Evaluating Counterfactual Prediction and Planning Error in a Compact World Action Model
pico-WAM: Evaluating Counterfactual Prediction and Planning Error in a Compact World Action Model
This paper introduces pico-WAM, a compact world action model designed to trace policy-dynamics interactions.
The study uses controlled interventions on a compact world action model trained on Push-T demonstrations to isolate the impact of counterfactual branch training and simulator-based scoring on policy search coverage, concluding that improving trajectory scoring is an effective target in this specific setting.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
demonstrations
206
View original evidence
arXiv:2611.01039v1 Announce Type: new 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-horiz
Open source S2training seeds
15
View original evidence
arXiv:2611.01039v1 Announce Type: new 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-horiz
Open source S2
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2611.01039v1 Announce Type: new 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-horiz
Open source S2
on 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 s
Open source S3
pico-WAM: Evaluating Counterfactual Prediction and Planning Error in a Compact World Action Model
Open source S1
Source：arXiv Robotics — research abstracts · arxiv.org