RuleGLM: An Ethics Evaluation Framework with Knowledge Vector Space

Existing robots cannot self-assess behavioral ethics, and it is difficult to constrain their actions according to ethics rules in the actual environment. To this end, this article innovatively proposes an ethics evaluation framework RuleGLM using multimodal large language model (MLLM) to conduct real-time ethics evaluation of robot behavior. Firstly, this article proposes a visual-reasoning module to realize the extraction and formatted output of environmental information. Then, this article encodes the ethics rules into a knowledge tensor base through the language-reasoning module. Finally, this article constructs the relationship mapping between rules in the knowledge tensor base through knowledge vector space and proposes a quantitative ethics evaluation algorithm to improve the interpretability and evaluation accuracy of RuleGLM. Experimental results show that compared with the iFlytekSpark model, the output format accuracy of RuleGLM increases from 44.17% to 98.61%. Compared with the baseline, the ethics evaluation algorithm integrated into the knowledge vector space has improved its accuracy of evaluation results from 38.99% to 89.94%. Data and code will be released at https://github.com/7Lin97/RuleGLM.