Збережено в:
| Автор: | |
|---|---|
| Формат: | Recurso digital |
| Мова: | |
| Опубліковано: |
Zenodo
2025
|
| Предмети: | |
| Онлайн доступ: | https://doi.org/10.5281/zenodo.17203749 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
Зміст:
- <p><p>This work introduces and proposes a novel, bio-inspired paradigm for robotic learning. We present the conceptual framework for a hierarchical system that combines a large Vision-Language Model (VLM) as a high-level semantic evaluator with a distributed, low-level "Cell Assembly" motor cortex. The core of our contribution is the detailed formulation of a Reinforcement Learning from AI Feedback (RLAIF) loop for complex robotic tasks, such as humanoid locomotion.</p></p> <p><p>The proposed Cell Assembly, composed of a large set of simple, independent binary-like cells, generates actions. A VLM (e.g., LLaVA-1.5 7B) would then visually observe the outcome of these actions and provide a qualitative reward signal based on an abstract goal (e.g., "stand upright"). This feedback would be used to update the parameters of the cellular network via a policy gradient algorithm, as detailed in our provided mathematical formulation and pseudocode.</p></p> <p><p>This research serves as a theoretical foundation for creating more autonomous and adaptive learning agents, shifting the paradigm from manually engineered rewards to AI-driven "education," and from monolithic architectures to robust, hardware-inspired cellular systems. The core scientific documents outlining this framework are available at the project's GitHub repository: https://github.com/zorino96/VLM-Robot-Director</p></p>