Rethink Before You Execute: Adaptive Execution for World Action Models

2026-08-10

AI Summary

This work addresses the inefficiency and reduced success rates of fixed-step World Action Models (WAMs) caused by their inability to adapt to varying reliability across task phases, which leads to poorly timed replanning. To overcome this limitation, the authors propose TempoWAM—a lightweight, plug-and-play adaptive execution framework that dynamically triggers replanning through a recurrent progress monitor evaluating real-time task advancement, coupled with an adaptive execution protocol. Additionally, task-specific calibration factors are introduced to align training and deployment conditions. By integrating task instructions, execution history, and online progress signals, TempoWAM significantly improves the efficiency–success trade-off across LIBERO, RoboTwin, and real-robot experiments: it reduces inference steps by 26.9% on simple tasks without compromising success rates, while boosting success by 13.3 percentage points on challenging tasks.

Abstract

World Action Models (WAMs) jointly predict future actions and the evolution of the environment. At each inference, a WAM generates a chunk of actions and the robot executes a fixed prefix before replanning. We argue that this fixed execution horizon is poorly matched to execution dynamics: the chunk reliability varies across task stages, so when to replan depends on the result of accumulated execution, not on the step counts. We propose TempoWAM (Timing Execution by Monitoring Progress Online), a lightweight plug-and-play execution scheme for WAMs. A Recurrent Progress Monitor first estimates task progress from the current observation, task instruction, remaining actions, and execution history; and an Adaptive Execution Protocol then evaluates whether the chunk is advancing the task to decide if replanning is needed. To bridge the training-deployment gap, the protocol is calibrated by a task-dependent calibration factor with online adaptation. Experiments on LIBERO, RoboTwin, and real-world tasks show that TempoWAM consistently improves the efficiency-success trade-off of WAM execution. On real robots, it reduces WAM inferences by 26.9% on easy tasks while maintaining success, and improves success by 13.3 points on difficult tasks.

Problem

• World Action Models
• adaptive execution
• replanning
• execution horizon
• task progress

Innovation

• World Action Models
• Adaptive Execution
• Progress Monitoring
• Online Calibration
• Replanning Strategy