Light-WAM: Efficient World Action Models with State-Fusion Action Decoding

2026-06-06

AI Summary

This work addresses the high training costs and substantial inference latency of existing World Action Models, which rely on large-scale generative architectures and hinder their deployment as efficient closed-loop policies. To overcome these limitations, the authors propose a lightweight World Action Model featuring a compact video backbone that learns from future video supervision in a downsampled latent space. The model incorporates a StateFusionActionExpert module to aggregate multi-layer state features and employs a query pooling mechanism to directly predict action chunks, thereby circumventing the need for heavyweight generative action experts. With only 0.44 billion trainable parameters, the model achieves strong performance on the LIBERO benchmark, demonstrates effective multitask capabilities on RoboTwin 2.0, and attains an inference latency of 72.03 ms with a peak GPU memory consumption of 4.1 GiB, significantly improving both training throughput and deployment efficiency.

Abstract

World Action Models (WAMs) extend robot policy learning by incorporating future prediction as an additional training objective, encouraging the policy to encode task-relevant temporal structure in its representations. Current WAMs often rely on large-scale generative architectures that incur high training costs and inference latency, making them difficult to deploy as efficient closed-loop policies. We propose Light-WAM, a lightweight World Action Model for efficient robot manipulation. Specifically, it is built with a compact video backbone and performs future-video supervision in a downsampled latent space, reducing the cost of video co-training while retaining its benefits for representation learning. For action prediction, Light-WAM introduces the StateFusionActionExpert, which reads adapted states from multiple backbone layers, fuses them through learned-query pooling, and directly predicts action chunks in a single forward pass. This design provides an efficient interface between video backbone representations and robot actions, avoiding the need for heavy generative action experts. Experiments demonstrate that Light-WAM maintains strong performance on LIBERO and achieves usable multi-task performance on RoboTwin 2.0, while using only 0.44B trainable parameters. It also achieves 72.03ms inference latency with 4.1GiB peak GPU memory and improved training throughput.

Problem

• World Action Models
• efficient robot policy
• high training cost
• inference latency
• closed-loop deployment

Innovation

• World Action Models
• State-Fusion Action Decoding
• Lightweight Robotics
• Latent Video Prediction
• Efficient Policy Learning