大语言模型策略适应下的随处驾驶（英）B-Li等-2024

Unique traffic code in current locationNew York City Driving HandbookScene DescriptionUnexpected Situation"Stoppedat a red light on the rightmost"someone honks at me"lane,starting to turnright"Traffic Rule ExtractorEnvironmentPlan(TRE)LLM PlannerNew Plan"Tum right"As per NYC traffic rues,even if you arein the rightmost lane,youshould notturn right on a red light;wait until thelight tums green to proceedsafely"Figure 2.Overview of LLaDA.In this illustration,the driver learned how to drive in California but now needs to drive in New York City.However,the road situation,traffic code,and unexpected situations are different.In our system,we consider three inputs:initial plan("Turn righr"),unique traffic code in current location (New York City Driving Handbook),and unexpected situation ("someone honks atmne").We will feed these three inputs into a Traffic Rule Extractor (TRE).which aims to organize and filter the inputs and feed the outputinto the frozen LLMs to obtain the final new plan.In this paper,we set GPT-4 as our default LLM.Collect Information"Continue driving safely,ignore the honk unless it's anbackground ='The condition:The driver get theemergency vehicle or you're causing obstruction."license from O.but now in 0 so we make sure thePlandifference in 0 and notify the drivers what they needNew Planto do.".format('California','Germany,'Germany)request ="0.0".format(plan,unexpected situation)response_format ='Please out put theinst ructionsUnexpectedLLM Plannerabout what should the driver do in one sentence lessSituation(GPT-4)than 20 words."Someone honks at me'prompt=background+'nThe driver says:'+request+n'+response_formatprompt￥Find KeywordsProcess Guidelines with Keywordsprompt 'Extract less than 3 common traffic-relatedphrases from the given sentence,each phraseSplit the content into paragraphscontains 1 or 2 words:0".format(prompt)paragraphs user_guidelines.split("nn')Search for keywords in each paragraphkeywords-for paragraph in paragraphsfor keyword in keywords:"Go straight""honk"LLMs(GPT-4)if keyword.lower0 in paragraph.lower0:Case-insensitive searchmatching_paragraphs append(paragraph)processed_guidelines=Find the paragraphs containing the keywordsUnique traffic codefor paragraph in matchingparagraphs:in current locationprocessed_guidelines +paragraph+'n'+50+1n'Germany Driving HandbookFigure 3.Details of Traffic Rule Extractor (TRE).As is shown in the figure,we first organize the information (such as locations,"Turnright"and "someone honks at me")into a prompt.Then we feed the prompt to find the one or two keywords using GPT-4.To guaranteethe search quality,each keyword contains one or two words.Then we find the key paragraphs that contain extracted keywords in the uniquetraffic code.In this way,we could filter out the necessary information and only organize the valuable material into GPT-4 to obtain thefinal new plan.rules relevant to the current scenario that the vehicle findswords in the traffic code of the current location,which