一种基于可扩展世界模型的机器人多任务持续学习方法
By constructing a modular world model that separates cognition and decision-making, and combining it with model predictive control, the problems of catastrophic forgetting and insufficient generalization in robot learning in multi-task and dynamic environments are solved, achieving efficient and stable learning results.
Patent Information
- Application Number
- CN202610884007.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-18
- Publication Date
- 2026-07-17
AI Technical Summary
Existing robots suffer from catastrophic forgetting, insufficient generalization ability, and low sample efficiency in multi-task and dynamic environments. Traditional world model methods are difficult to adapt to multi-environment and multi-task scenarios, resulting in degraded model performance and low learning efficiency.
It adopts a modular architecture based on a scalable world model, including a cognitive layer and a decision layer. The cognitive layer performs unified modeling of the environment and tasks, while the decision layer performs action sequence prediction and reward evaluation through multiple reusable skill modules. Combined with model predictive control, it achieves efficient learning.
It effectively alleviates the problem of catastrophic forgetting, improves the adaptability and generalization ability to new tasks, increases sample utilization efficiency, reduces training costs, and maintains the stability and compactness of the model.
Smart Images
Figure CN122401455A_ABST