Robotics · 2026-10-30
Utility-WAM: Correcting Future-Induced Action Changes
Feihong Zhang, Xi Lin, Yulong Shi, Zuxing Lu, Yanghong Mei, Zhaopeng Gu, Tingyu Yuan, Zihao Liang, Yinuo Wang, Zhaowen Li
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Utility-WAM studies how predicted visual futures change robot actions via an action delta between future-conditioned and future-free execution. Authors report futures barely change average success (+0.8 LIBERO, −0.8 RoboTwin 2.0) but increase inference latency (1.7–1.9×). They introduce a residual adapter plus a pre-generation gate to correct future-induced action changes and skip video when unhelpful.
Why it matters
The paper identifies a concentrated failure mode of using test-time imagination: future-induced negative transfer that harms precision tasks while sometimes helping long-horizon tasks. Utility-WAM targets the harmful action changes rather than disabling futures globally, and also reduces inference cost with selective future generation.
Method
- Characterizes test-time imagination via the action delta Δa between future-conditioned and future-free action execution; future-induced negative transfer is measured as closed-loop success change relative to the future-free path.
- Introduces Utility-WAM: a residual adapter trained on a frozen shared WAM to repair future-induced velocity deltas using ground-truth dynamic guidance alignment and an excess-error penalty beyond the future-free baseline.
- Trains a lightweight pre-generation gate from adapter-paired outcomes to choose between the corrected future-conditioned path and the future-free path, skipping video generation when correction is not useful.
Limitation
The provided text does not state a specific stated limitation.
Abstract (from arXiv)
World Action Models (WAMs) can condition robot actions on generated visual futures, but this additional context is neither cost-free nor uniformly beneficial. We examine what a predicted future actually changes: the action delta $\Delta a$ between the future-conditioned and future-free execution paths of the same model. Under a shared WAM checkpoint, this delta barely shifts average benchmark success ($+0.8$ points on LIBERO, $-0.8$ on RoboTwin~2.0), while generating visual futures increases inference latency by 1.7--1.9$\times$. These aggregate metrics conceal a concentrated failure mode that we term \emph{future-induced negative transfer}: while predictive deltas benefit long-horizon tasks ($+2.4$ points on LIBERO-Long) and remain negligible on most routine manipulations, they induce severe degradation on precision tasks---with \emph{Press Stapler} and \emph{Place Can Basket} losing 54.0 and 34.0 points, recurring as top failure modes across independently trained WAM variants. We introduce \textbf{Utility-WAM} for in-place correction of future-conditioned control. Its primary learning component is a residual adapter that repairs future-induced velocity deltas under a frozen shared WAM by aligning them with ground-truth dynamic guidance and penalizing excess error beyond the future-free baseline. The adapter selectively eliminates the majority of severe task drops (reducing losses to 6.0 and 14.0 points) while preserving long-horizon foresight. To eliminate superfluous video computation, a lightweight pre-generation gate reuses counterfactual action outcomes to selectively bypass video generation on 75.5\% of RoboTwin queries, cutting latency by 38\% while improving average success to 93.3\% on RoboTwin~2.0 and 98.9\% on LIBERO. On physical bimanual manipulation across three real-world tasks, Utility-WAM succeeds in 53 of 60 trials, compared to 43 for FAST-WAM.
Related papers
- Achieve What You Imagined: Learning to Align Actions with Visual Plans
- Plausible Future, Wrong Choice: Steering World-Action Models toward Intended Future
- WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction
- What Limits Cross-Task Generalization in World Action Models?
- {\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: 2026-10-29 · Indexed by Pub-AI: 2026-10-30 UTC
Utility-WAM: Correcting Future-Induced Action Changes
Preprint posted: 2026-10-29 · Original record ↗
- arxiv
- 2610.39367
Explicit source versions (1)
- v1 · 2026-10-29
- Preprint available
Source date: 2026-10-29 · First detected: 2026-10-30 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.39367v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
World Action Models (WAMs) can condition robot actions on generated visual futures, but this additional context is neither cost-free nor uniformly beneficial. We examine what a predicted future actually changes: the action delta $Δa$ between the future-conditioned and future-free execution paths of the same model. Under a shared WAM checkpoint, this delta barely shifts average benchmark success ($+0.8$ points on LIBERO, $-0.8$ on RoboTwin~2.0), while generating visual futures increases inference latency by 1.7--1.9$\times$. These aggregate metrics conceal a concentrated failure mode that we term \emph{future-induced negative transfer}: while predictive deltas benefit long-horizon tasks ($+2.4$ points on LIBERO-Long) and remain negligible on most routine manipulations, they induce severe degradation on precision tasks---with \emph{Press Stapler} and \emph{Place Can Basket} losing 54.0 and 34.0 points, recurring as top failure modes across independently trained WAM variants. We introduce \textbf{Utility-WAM} for in-place correction of future-conditioned control. Its primary learning component is a residual adapter that repairs future-induced velocity deltas under a frozen shared WAM by aligning them with ground-truth dynamic guidance and penalizing excess error beyond the future-free baseline. The adapter selectively eliminates the majority of severe task drops (reducing losses to 6.0 and 14.0 points) while preserving long-horizon foresight. To eliminate superfluous video computation, a lightweight pre-generation gate reuses counterfactual action outcomes to selectively bypass video generation on 75.5\% of RoboTwin queries, cutting latency by 38\% while improving average success to 93.3\% on RoboTwin~2.0 and 98.9\% on LIBERO. On physical bimanual manipulation across three real-world tasks, Utility-WAM succeeds in 53 of 60 trials, compared to 43 for FAST-WAM.