GLOW: Omni-Intelligence World Model for Lunar and Martian Exploration

A embodied general intelligence world model co-developed by CSUN and NASA JPL, integrating generative AI and embodied learning. It supports autonomous rover navigation, robotic manipulation, and multi-robot coordination, providing predictive decision-making capabilities for planetary exploration missions.

• Original Authors/Maintainers: GLOW Community (CSUN Computer Science / Embodied AGI, NASA JPL)
• Source Platform: GitHub
• Original Title: GLOW: Generative AI and Embodied Learning for Omni-Intelligence World Model
• Original Link: https://github.com/GLOW-Community/GLOW
• Project Homepage: https://glow-community.github.io/GLOW/
• Release Date: June 2026

NASA's CADRE (Cooperative Autonomous Distributed Robotic Exploration) mission plans to deploy a cluster of collaborative rovers on the lunar surface. However, unstructured terrain, communication delays, occlusions, and uneven surfaces on the Moon and Mars pose significant challenges to autonomous navigation. Traditional reactive strategies, which directly map observations to actions, often fail in unknown environments.

The GLOW project was born to build an Omni-Intelligence World Model that supports predictive decision-making by forecasting future environmental states, rather than just reacting to current observations. This capability is crucial for autonomous systems performing tasks in extreme environments like the Moon and Mars.

The core of GLOW is a world model engine that can predict future environmental states from current observations and actions. Unlike reactive strategies, the world model (e.g., BAGEL 14B integrated in π0.7) can forecast future observations, which is critical for navigating unstructured terrain with occlusions and uneven surfaces.

The project combines generative AI and embodied learning, enabling robots to acquire visuomotor skills through real-world and simulated experiences, and possess strong sim-to-real transfer capabilities. This integration allows the system to generalize to new scenarios without retraining for specific tasks.

GLOW supports multi-robot coordination within an autonomous AGI modeling framework— a key capability for NASA's next-generation planetary exploration missions. Through learned reasoning and reasoning representations, rovers can make joint decisions without explicit centralized control.

The GLOW architecture is built around three operational pillars:

Operation & Execution

Responsible for high-level strategy generation, sub-goal planning, and language instruction parsing. This pillar converts abstract task objectives into concrete action sequences executable by robotic agents. Task planning is driven by contextual learning from large robotic foundation models (e.g., π0.7), enabling the system to generalize to new scenarios.

Trustworthy Inference & Interpretability

Aligned with NASA's gold mission standards for safety, ensuring that reasoning decisions are interpretable, verifiable, and based on knowledge reasoning—requirements for mission-critical autonomous systems. This pillar is essential for establishing system credibility in deep space missions where human intervention is impossible.

Spatial Intelligence

Optimizes actions through spatial understanding, video prediction, and environmental modeling. This pillar enables the system to build rich 3D environmental representations for planning and control. Robots can "imagine" future scenarios in their minds and select optimal action sequences.

GLOW is built on existing large foundation models, including:

• π0.5/π0.7: Hierarchical vision-language-action architecture with world model
• Gemini: Google's multimodal large model
• GPT-5: OpenAI's general-purpose large language model

This integration strategy allows GLOW to leverage the general capabilities of pre-trained models and focus on robot-specific world modeling and decision-making.