Source-linked AI summary
Completion Aware Guidance for World Action Models
Seungyeon Kim, Junhoo Lee, Baekseung Kim, Minkyu Kim, Nojun Kwak
TL;DR
WAMs can generate plausible, action-consistent futures that omit transitions needed for task completion, and this failure is associated with adaptation to short-horizon control. The paper introduces training-free Completion Aware Guidance to steer sampling toward phase-completing transitions. Across representative WAM evaluations, it reports higher task success and less task-incomplete imagination.
Problem
WAMs can produce visually plausible, action-consistent predictions that omit transitions required for task completion.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
Using the same backbone, native-horizon generation produces the completion event while short-chunk autoregressive generation repeatedly omits it, suggesting the failure is not intrinsic to the backbone.
Source: §3 · p. 2CAG is currently limited to sampling-time intervention; extending it to WAM training is left for future work.
Source: §5 · p. 4
Abstract
from arXiv · showhide
World Action Models (WAMs) predict visual futures and robot actions, yet they remain susceptible to task-incomplete imagination, where plausible, action-consistent predictions omit the transition needed for task completion. In this paper, we show that this failure is not inherent to the world model backbone, but emerges when adapted for short-chunk control, which can repeatedly favor plausible local continuations over task-completing transitions. To address this, we introduce Completion Aware Guidance (CAG), a training-free sampling method that guides generation toward task completion. Across representative WAMs, CAG improves success from 64% to 70% on a RoboTwin 2.0 subset and from 69% to 75% in zero-shot simulation, while reducing task-incomplete imagination from 79% to 40%.
1 Introduction
WAMs can imagine coherent, action-consistent futures that omit transitions needed to complete a task. The paper attributes this to a mismatch between fixed-horizon video pretraining and short-horizon WAM control, and introduces training-free Completion Aware Guidance (CAG).
1 Introduction
- Task-incomplete imagination produces plausible, instruction-relevant video and actions that agree but omit a task-completing transition, such as releasing an object above its target.The predicted actions follow the incomplete future, resulting in task failure.Source: §1 · p. 1
- The paper attributes this failure to a mismatch between fixed-horizon video-generation pretraining and short-horizon WAM control, which can repeatedly defer the required transition.WAM chunks are executed and regenerated from observation, and each horizon need not complete the current phase.Source: §1 · p. 1
- CAG is a training-free sampling method that strengthens instruction conditioning relevant to the current video-action prediction across DreamZero and Fast-WAM [6] [7].It steers the current chunk toward the required state transition without retraining the WAM.Source: §1 · p. 1
- Task success rises from 64.4% to 70.0% and from 69% to 75%, while task-incomplete imagination falls from 79% to 40%.The success-rate pairs correspond to a RoboTwin 2.0 subset and zero-shot simulation, respectively; the failure reduction is reported in a representative manipulation setting.Source: §1 · p. 2
2 Related Work
Prior work develops vision-language-action policies and world action models for robot control, alongside approaches that make task progress more explicit. The paper distinguishes specifying an intermediate target from producing the transition that reaches it, and identifies a gap in prior WAM analyses.
2 Related Work
- Vision-language-action policies map visual observations and language instructions to actions or action chunks [1] [2] [3] [4] [5].For temporally extended tasks, a global instruction may not specify the next semantic subtask or desired near-future state.Source: §2 · p. 2
- Some approaches provide language subtask instructions [9] or visual subgoal images [10] [11], but specifying an intermediate target does not ensure the transition that completes it.The distinction is between defining a subtask goal and reaching it.Source: §2 · p. 2
- WAMs couple imagined visual trajectories with actions as implicit temporal plans [6] [7] [21] [22] [8], while prior work [6, 7, 23, 24] does not isolate plausible, mutually consistent continuations that omit a phase-completing transition.The paper names this failure task-incomplete imagination and mitigates it with training-free sampling guidance.Source: §2 · p. 2
3 Method
Task-incomplete imagination is a plausible video-action continuation that omits the transition needed for phase completion. The section contrasts short-chunk generation with longer-horizon generation and presents CAG as a way to guide sampling toward completion.
3 Method
- The same video diffusion backbone produces task-completing transitions at its extended 5-second horizon, while short-chunk autoregressive generation repeatedly omits them.The comparison suggests task-incomplete imagination arises from adaptation to short-horizon control rather than an intrinsic backbone limitation.Source: §3 · p. 2, Figure 2 · p. 3
- CAG treats task-incomplete imagination as a selection problem, biasing generation toward a continuation that completes the active manipulation phase while staying close to the WAM prior.Short-horizon generation can repeatedly defer a required transition in favor of a locally plausible holding or hovering continuation.Source: §3 · pp. 2–3
- The ideal guided distribution favors chunks that complete the active phase while a KL regularizer keeps it close to the original WAM prior.The phase-completion indicator marks whether a predicted chunk contains the transition completing the active phase.Source: §3 · p. 3
- The completion sensitivity vector identifies which instruction components would increase completion elicitation, but it is unavailable during sampling, motivating use of observable attention signals.Source: §3 · pp. 3–4
- CAG uses cross-attention to identify instruction components already used by the current video-action generation, then strengthens selected components to bias the next chunk toward completion.The intervention is restricted to the instruction-token key and value pathway; a sparse gate controls which signals are strengthened.Source: §3 · pp. 3–4
4 Experiments
CAG improves task success across the evaluated WAMs and reduces task-incomplete imagination in DreamZero failure analysis.
Setting
- Evaluation used 50 rollouts per official DreamZero DROID simulation scene and 30 per each of nine non-saturated RoboTwin 2.0 tasks.The RoboTwin tasks had published Fast-WAM Random success rates below 90%; α = 0.15 was fixed across tasks and backbones without per-task tuning.Source: §4 · p. 4
Results
- 64.4% to 70.0% average task success on nine RoboTwin 2.0 tasks, and 69% to 75% on three DreamZero zero-shot tasks.CAG gains were reported on every DreamZero task.Source: §4 · p. 4
Failure Analysis
- 79% to 40% of DreamZero zero-shot failures were attributable to task-incomplete imagination with CAG.The analysis examines failure cases to assess whether CAG addresses this failure mode.Source: §4 · p. 4
5 Conclusion
The paper identifies task-incomplete imagination and attributes it to a mismatch between fixed-horizon video generation and short-horizon WAM control. CAG addresses this through sampling-time guidance, while extending it to WAM training remains future work.
5 Conclusion
- Task-incomplete imagination is a plausible, action-consistent continuation that omits the transition needed to complete the current task phase.The paper attributes this failure to the mismatch between fixed-horizon video generation and short-horizon WAM control.Source: §5 · p. 4
- CAG strengthens instruction conditioning relevant to the current prediction to elicit task-completing transitions without retraining.Across representative WAMs, it improves task success and reduces task-incomplete imagination.Source: §5 · p. 4
- CAG is currently limited to sampling-time intervention; extending it to WAM training is future work.Source: §5 · p. 4