Learning Real-World Action-Video Dynamics with Heterogeneous Masked Autoregression

TL;DR

This paper introduces Heterogeneous Masked Autoregression (HMA), a novel model for generating high-quality, controllable action-video data for robotics, achieving faster and more realistic video synthesis across diverse robotic settings.

Contribution

HMA is the first to leverage heterogeneous pre-training and masked autoregression for scalable, real-time robotic video generation across multiple domains and tasks.

Findings

HMA outperforms previous models in visual fidelity and controllability.

HMA achieves 15x faster real-world video generation.

Post-trained HMA effectively evaluates policies and generates synthetic data.

Abstract

We propose Heterogeneous Masked Autoregression (HMA) for modeling action-video dynamics to generate high-quality data and evaluation in scaling robot learning. Building interactive video world models and policies for robotics is difficult due to the challenge of handling diverse settings while maintaining computational efficiency to run in real time. HMA uses heterogeneous pre-training from observations and action sequences across different robotic embodiments, domains, and tasks. HMA uses masked autoregression to generate quantized or soft tokens for video predictions. \ourshort achieves better visual fidelity and controllability than the previous robotic video generation models with 15 times faster speed in the real world. After post-training, this model can be used as a video simulator from low-level action inputs for evaluating policies and generating synthetic data. See this link https://liruiw.github.io/hma for more information.