QuantWAMs: Calibrating at the Right Granularity for World Action Models

2026-07-30

AI Summary

This work addresses the challenge of efficiently deploying world action models (WAMs) with existing post-training quantization methods, which often rely on open-loop objectives, homogeneous assumptions, and calibration distributions misaligned with real-world deployment conditions. To overcome these limitations, we propose QuantWAMs—the first post-training quantization framework tailored for WAMs—leveraging a fine-grained calibration strategy that integrates model architecture, trajectory distribution, and task-specific objectives. Key innovations include shared-base outlier calibration, joint objective-aware saliency weighting, fixed-intervention trajectory auditing, empirical Fisher-based saliency analysis, coordinate-compatible activation pooling, and closed-loop reachable-state-guided denoising scheduling. Under W4A4 quantization, our method incurs only a 0.2–0.7 percentage point performance drop in simulated tasks and successfully executes three real-world robotic manipulation tasks, achieving a peak memory footprint of 29% relative to FP16 and module-level speedups of 1.4–1.6×.

Abstract

World Action Models (WAMs) jointly predict future observations and actions, but their iterative denoising and closed-loop execution make efficient deployment costly. Existing post-training quantization (PTQ) methods are poorly suited to WAMs because they rely on open-loop objectives, homogeneous model assumptions, and calibration distributions that do not reflect deployment. We present QuantWAMs, a PTQ framework that aligns quantization decisions with the calibration context defined by model structure, rollout distribution, and task objective. QuantWAMs introduces three strategies: shared-basis outlier calibration, which pools activation evidence only across coordinate-compatible modules; co-training-objective saliency, which computes empirical-Fisher scores from the joint video--action gradient and assigns weight precision at a calibration-stable layer granularity; and fixed-intervention rollout auditing, which revises denoising-step protection schedules using reachable closed-loop states without changing the precision budget. We evaluate QuantWAMs on Fast-WAM and LingBot-VA across RoboTwin 2.0, LIBERO, and real-robot manipulation with an AgiBot G2. Under a W4A4-dominant setting, the reported simulation means differ from FP16 by 0.2--0.7 percentage points. Real-robot trials further establish deployment feasibility on three manipulation tasks. For the targeted video and action blocks, QuantWAMs reduces peak weight-and-activation memory to about 29\% of FP16 and provides 1.4--1.6$\times$ block-level speedups.

Problem

• World Action Models
• post-training quantization
• closed-loop execution
• calibration distribution
• efficient deployment

Innovation

• QuantWAMs
• post-training quantization
• World Action Models
• closed-loop deployment
• calibration granularity