Robot foundation models are large-scale AI models trained on diverse robotic data that enable robots to understand and perform a wide variety of tasks without task-specific training. These models learn generalizable representations of actions, objects, and environments that transfer across different robot platforms, tasks, and scenarios. Cross-embodiment capability means that knowledge learned on one type of robot (e.g., a robotic arm) can transfer to different robot forms (e.g., a humanoid or mobile robot), dramatically reducing the training required for each new application.
The foundation model approach mirrors the success of large language models in natural language processing, applying similar principles to robotics. By training on massive datasets of robotic demonstrations, sensor data, and task executions, these models develop a general understanding of manipulation, navigation, and interaction that can be adapted to specific tasks with minimal additional training. This enables robots to quickly learn new tasks, adapt to novel situations, and operate in diverse environments. The technology is fundamental to developing general-purpose robots that can perform many different tasks rather than being specialized for single applications, representing a major step toward truly versatile robotic systems.
AI robotics company building a universal AI brain for robots.
Developers of the Gemini family of models, which are trained from the start to be multimodal across text, images, video, and audio.
A startup building a general-purpose brain for robots, backed by OpenAI and Thrive Capital.
Developing foundation models for robotics (Project GR00T) and vision-language models like VILA.
Building a shared general-purpose brain for diverse robot embodiments, leveraging massive training data.
A Norwegian robotics company (backed by OpenAI) developing androids like EVE and NEO.
Developing general-purpose humanoid robots designed for commercial workforce deployment.
The global hub for open-source AI models and datasets. Founded by French entrepreneurs with a major office in Paris.
Developing general-purpose humanoid robots (Phoenix) powered by Carbon, their AI control system.
R&D arm of Toyota Motor Corporation.
Paper
GR00T N1: An Open Foundation Model for Generalist Humanoid RobotsarXiv · Mar 12, 2025
Introduction of GR00T N1, a Vision-Language-Action (VLA) foundation model with a dual-system architecture designed for generalist humanoid robots, capable of interpreting environments and generating motor actions.
Paper
HALO: A Unified Vision-Language-Action Model for Embodied Multimodal Chain-of-Thought ReasoningarXiv · Feb 1, 2026
Proposal of HALO, a unified VLA model enabling embodied multimodal chain-of-thought reasoning, using a Mixture-of-Transformers architecture to decouple reasoning, foresight, and action prediction.
Paper
OmniVLA: An Omni-Modal Vision-Language-Action Model for Robot NavigationarXiv · Sep 1, 2025
Presentation of OmniVLA, a foundation model trained on over 9,500 hours of data across 10 platforms, capable of processing omni-modal goals (images, poses, language) for robust navigation.
Paper
ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot ControlarXiv · Feb 1, 2026
ZEST introduces a framework for zero-shot transfer of athletic skills to humanoid robots using reinforcement learning trained on diverse motion sources, deployed on Boston Dynamics' Atlas.
