DreamZero

14B World-Action Model (WAM) for robotic manipulation via video diffusion on EKS

View on GitHub

Overview

DreamZero is a 14B-parameter World-Action Model (WAM) — a Wan-based, causal (autoregressive) video diffusion transformer that jointly denoises future video frames and action vectors for robotic manipulation.

This test case deploys DreamZero LIBERO 14B SFT training on Amazon EKS.

Key Features

  • 14B parameter video diffusion transformer
  • Joint video + action prediction (World-Action Model)
  • Causal autoregressive generation
  • LIBERO manipulation benchmark
  • EKS deployment with GPU scheduling

⚠️ Non-commercial license. The DreamZero-DROID model uses CC-BY-NC-4.0. The code in the repository is MIT-0.