Policy distillation method, computer device and computer storage medium

By introducing the noise intensity dynamically generated by policy entropy and the relabeling technique of the coach model into the policy distillation method, the problems of noise adjustment dependence and low data efficiency of existing policy distillation methods are solved, achieving adaptive training and efficient generalization, which is suitable for high data acquisition cost scenarios such as robotics.

CN122334400APending Publication Date: 2026-07-03SHENZHEN INST OF ARTIFICIAL INTELLIGENCE & ROBOTICS FOR SOC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610465625.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-09
Publication Date
2026-07-03

AI Technical Summary

Technical Problem

Existing strategy distillation methods suffer from low data utilization efficiency due to reliance on manual tuning for noise adjustment and poor scalability in multi-tasking environments, making them difficult to apply effectively in high-data-acquisition-cost scenarios such as real robots.

Method used

By introducing noise intensity dynamically generated based on policy entropy and relabeling training trajectory data using a trained coaching model, optimal action labels are generated, optimizing the student policy training process and achieving adaptive control of noise parameters and construction of high-quality training samples.

Benefits of technology

It avoids the tediousness of manually adjusting noise and the problem of hyperparameter sensitivity, reduces data sampling costs and computational resource consumption, improves the ease of use and versatility of the method, and enhances the model's generalization ability and robustness in multi-task environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122334400A_ABST
    Figure CN122334400A_ABST
Patent Text Reader

Abstract

This application discloses a policy distillation method, a computer device, and a computer storage medium. By introducing noise intensity dynamically generated based on policy entropy and using a trained coach model to recalculate the optimal action labels from the states in the training trajectory data, adaptive control of noise parameters and construction of high-quality training samples are achieved. This approach avoids the tediousness of manually adjusting noise and the hyperparameter sensitivity issues of the DPD method, eliminating the need for repeated trial-and-error optimization for different tasks, thus improving the ease of use and versatility of the method. Furthermore, unlike OPD, it does not require real-time interaction with the environment during training to obtain feedback. Instead, it significantly reduces data sampling costs and computational resource consumption by combining offline training trajectory data relabeling with dynamic noise modulation, overcoming efficiency bottlenecks and making it easier to implement in high-data-acquisition-cost scenarios such as real-world robotics.
Need to check novelty before this filing date? Find Prior Art