A recent arXiv paper introduces EmbodiedEvo, a coding agent designed for physical intelligence. Rather than relying on pre-trained policies or task-specific fine-tuning, the agent constructs executable programs and revises them based on feedback. This treats robot control as a code-generation problem, where the agent iteratively improves its own behavior.
The agent operates through a programmatic robot interface, observing the scene and generating code that maps observations to actions. Because the programs are executable and revisable, the system can adapt without needing new training data for every task. The authors position this as a foundation for embodied control that could generalize across different scenarios.
The abstract is brief, so details about the interface, evaluation, or comparisons to other methods are not yet available. As a single source, there are no conflicting claims to reconcile, but the approach's practical limits remain to be seen.