OpenWAM-Official/OpenWAM

What it solves

OpenWAM provides a modular research stack for developing World-Action Models (WAMs), which are models that combine world knowledge (predicting future states/video) and action learning (predicting robot actions). It aims to move away from tightly coupled design choices, allowing researchers to systematically experiment with different model architectures, representations, and training strategies to improve how robots inherit world knowledge and scale across domains.

How it works

OpenWAM is organized into three main components:

• OpenWAM-Infra: A modular infrastructure that allows users to compose and compare different video backbones (e.g., Wan, Cosmos), VLM backbones (e.g., Qwen3-VL), and WAM architectures (single, dual, or tri-system).
• OpenWAM-Study: A set of controlled studies used to derive principles for coupling world and action learning.
• OpenWAM-α: A large-scale pretrained foundation model trained on over 518 million frames of egocentric human and robot data.

Architectures range from simple shared systems to complex dual-systems with separate action and video DiTs, and tri-systems that incorporate a frozen VLM for high-level understanding.

Who it’s for

It is designed for AI researchers and robotics engineers working on world models and generalist robot policies, specifically those looking to pretrain and fine-tune models that can predict both visual outcomes and robot actions.

Highlights

• Modular Design: Supports interchangeable video backbones, visual encoders, and VLM backbones.
• Extensive Benchmarks: Integrated support for evaluation on RoboTwin, LIBERO, RoboCasa365, VLABench, and more.
• ** uma Foundation Model**: Includes OpenWAM-α, a pretrained model trained on 6,400 hours of data.
• Flexible Architectures: Offers multiple system configurations (single, dual, tri) to test how action and world knowledge are integrated.