UnifoLM-WMA-0: A World-Model-Action (WMA) Framework Under UnifoLM Family (unigen-x.github.io)

Unitree released UnifoLM-WMA-0, an open-source “world-model–action” (WMA) framework for general-purpose robot learning that couples a learned world model with action heads across multiple robot embodiments. The world model serves two tightly integrated roles: (1) a Simulation Engine that generates high-fidelity, interactive synthetic rollouts from a current image plus planned future actions, and (2) a Policy Enhancement/Decision-Making mode that predicts future physical interactions to inform and optimize action selection. The team fine-tuned a video-generation backbone on the Open-X dataset and trained on five Unitree open datasets; results demonstrate action-controllable short-horizon generation and long-horizon interactive video prediction, and the system has been validated on real-robot deployments. For the AI/ML community this is significant because it operationalizes a model-based approach that unifies data synthesis and model-predictive-style policy improvement in a single architecture. Key technical implications: the framework produces task-conditioned synthetic trajectories (image + text/instruction inputs → future video), enabling scalable offline data augmentation and planning-aware policies; it supports multi-step, controllable generation based on explicit future action sequences; and it’s designed to bridge sim-to-real by training in a world-model loop and deploying the learned policy. Being open-source, UnifoLM-WMA-0 offers a reusable baseline for research in model-based control, video-conditioned policy learning, and long-horizon interactive prediction.