Robotics · 2026-10-06
$R^2$-WAM: Repair-and-Reject Post-Training for World Action Models
Ruiyan Xu, Haisheng Su, Sixu Lin, Zhaokun Yue, Chengming Hu, Xin Jin, Guiliang Liu
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
R2-WAM is a two-stage repair-and-reject post-training method for world action models. Predictive repair uses a kinematic alignment score to make imagined futures consistent with the input demonstrated action, then action rejection compares imagined progress of sampled vs demonstrated actions using STEAM to select inferior samples for negative fine-tuning. It improves downstream policy success without extra environment interaction or inference changes.
Why it matters
World Action Models can generate visually plausible but action-inconsistent futures, which can mislead consequence-based policy refinement. R2-WAM aims to fix this mismatch by (1) repairing action fidelity in the video expert via motion agreement with observed robot behavior, and (2) using the repaired model to perform selective negative fine-tuning based on imagined task progress.
Method
- Predictive repair: for each demonstration, use a kinematic alignment/action-following score measuring agreement between predicted and demonstrated robot motion to identify action-inconsistent generated futures and apply negative fine-tuning weights to them.
- Action rejection: freeze the repaired video expert, predict futures for sampled vs demonstrated actions from the same initial observation, score predicted task progress with a frozen temporal progress model following STEAM, and apply negative fine-tuning to sampled actions whose predicted progress falls below the demonstrated reference by a margin.
- Training/post-training happens without additional environment interaction and without changes to the inference procedure.
Limitation
The pointwise population analysis of the rejection objective assumes a fixed reference predictor and a fixed distribution of selected negatives; it “does not establish convexity in neural network parameters or characterize training dynamics under repeated negative sampling.”
Abstract (from arXiv)
World Action Models (WAMs) emerge as a promising foundation for policy refinement by predicting the consequences of sampled actions. However, visually plausible predictions can mislead policy refinement if they fail to reflect the input actions. To address this mismatch, we introduce $R^2$-WAM, a two-stage repair-and-reject post-training framework that first improves the consistency of predicted futures with input actions, then uses these futures to select inferior action samples for negative fine-tuning. The repair stage grounds imagination in observed robot behavior through a kinematic alignment score that measures agreement between predicted and demonstrated motion, enabling the predicted video to faithfully reflect its input actions. Using the repaired video model, the rejection stage compares imagined outcomes of sampled and demonstrated actions, selectively applying negative fine-tuning to samples whose predicted task progress falls below the demonstrated reference by a prescribed margin. Together, the two stages extend video prediction from representation learning to consequence-based policy refinement without additional environment interaction or changes to the inference procedure. $R^2$-WAM achieves 93.8% average success on RoboTwin 2.0 across clean and randomized settings. On the long-horizon real-world Fold Shirt task, it achieves 87.5% average success, compared with 0% for Fast-WAM.
Related papers
- EVO-WAM: Evolving World Action Models through Video-Action Verification
- Achieve What You Imagined: Learning to Align Actions with Visual Plans
- RAP-WAM: Learning World-Action Models through Predictive Rehearsal and Action Perturbations
- De-Diffusing WAMs: Action-Sufficient Direct Prediction for Robot Control
- {\Delta}WAM: Distilling Action Tangent Fields into World Action Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-04 · Indexed by Pub-AI: 2026-10-06 UTC
$R^2$-WAM: Repair-and-Reject Post-Training for World Action Models
Preprint posted: 2026-10-04 · Original record ↗
- arxiv
- 2610.04913
Explicit source versions (1)
- v2 · 2026-10-07
- Preprint available
Source date: 2026-10-04 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
- Explicit source version detected · v2
Source date: 2026-10-07 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
Abstract · arxiv
World Action Models (WAMs) emerge as a promising foundation for policy refinement by predicting the consequences of sampled actions. However, visually plausible predictions can mislead policy refinement if they fail to reflect the input actions. To address this mismatch, we introduce $R^2$-WAM, a two-stage repair-and-reject post-training framework that first improves the consistency of predicted futures with input actions, then uses these futures to select inferior action samples for negative fine-tuning. The repair stage grounds imagination in observed robot behavior through a kinematic alignment score that measures agreement between predicted and demonstrated motion, enabling the predicted video to faithfully reflect its input actions. Using the repaired video model, the rejection stage compares imagined outcomes of sampled and demonstrated actions, selectively applying negative fine-tuning to samples whose predicted task progress falls below the demonstrated reference by a prescribed margin. Together, the two stages extend video prediction from representation learning to consequence-based policy refinement without additional environment interaction or changes to the inference procedure. $R^2$-WAM achieves 93.8% average success on RoboTwin 2.0 across clean and randomized settings. On the long-horizon real-world Fold Shirt task, it achieves 87.5% average success, compared with 0% for Fast-WAM. Our project page is available at https://r2-wam.github.io/.