Toward Zero Fixed Code: Complete-Generation-Information-Driven Self-Learning for Autonomous Robots via Multi-Layer Code Replaceability and Safe Rollback

Autonomous robots operating in complex real-world environments require the ability to continuously improve their behavior beyond what can be fully specified by predefined code. However, conventional robotic systems usually rely on fixed hand-crafted logic, at least for some core components, which makes them difficult to adapt when new task constraints, environmental changes, or previously unseen problems arise after deployment. In this paper, we propose a complete-generation-information-driven self-learning framework for autonomous robots, in which the robot preserves not only the currently executable code but also the full information chain that produced it, including generation prompts, task objectives, execution traces, environmental feedback, and historical logs. Based on this preserved information, the robot can interact with a large language model to revise its own executable logic when new requirements appear. The proposed framework further supports multi-layer code replaceability, allowing not only task-solving code but also higher-level operational logic, such as update policy and scheduler logic, to be replaced during runtime. To maintain survivability during self-evolution, we also introduce a redundant replacement and rollback mechanism that enables the system to reject faulty candidate code and recover a previously valid version without interrupting continued operation. Experimental results show that on an 11-episode progressively changing task benchmark, the proposed method achieved a final cumulative success rate of 1.0, compared with 0.2727 for the fixed-code baseline, while completing 7 update events, 11 successful layer replacement events, and 1 rollback event.