Code For Paper Can LLM be a Good Path Planner based on Prompt Engineering? Mitigating the Hallucination for Path Planning

Comparison of the performance of S2RCQL and Rememberer

Comparison of the effectiveness of the S2RCQL algorithm without course or S2R and under different curriculum generation schemes.

The results for each model are in all mazes. The best results are highlighted in Bold, and the best baseline models are underlined. The score in each cell is represented as "ChatGLM-6B[ChatGPT score, ERNIE-Bot score]".

Method	Success (%) (5×5)	Optimality (%) (5×5)	Success (%) (7×7)	Optimality (%) (7×7)	Success (%) (10×10)	Optimality (%) (10×10)	Success (%) (100×100)	Optimality (%) (100×100)
Naive	9.2 [11.0, 11.4]	9.5 [11.5, 10.8]	8.5 [10.6, 10.3]	9.3 [13.0, 12.5]	7.4 [9.0, 9.1]	8.5 [9.2, 8.9]	-	-
CoT	12.5 [15.0, 15.0]	13.5 [15.4, 14.5]	11.1 [14.2, 14.9]	12.8 [14.2, 13.5]	8.9 [10.3, 10.5]	9.7 [11.7, 10.1]	-	-
ToT	13.7 [16.8, 17.1]	14.3 [14.6, 13.8]	12.0 [16.5, 16.6]	13.7 [14.4, 13.1]	9.0 [10.0, 10.3]	11.1 [12.5, 12.9]	-	-
ReAct	15.0 [17.6, 17.4]	15.8 [23.5, 22.8]	14.3 [17.4, 16.1]	15.0 [21.6, 21.6]	12.6 [14.8, 15.4]	14.4 [21.3, 20.7]	-	-
Remember	37.6 [44.7, 45.1]	46.4 [52.6, 50.8]	34.8 [37.4, 44.2]	37.8 [46.4, 40.2]	32.5 [32.6, 34.8]	30.5 [36.6, 35.7]	20.0 [26.7, 26.7]	50.0 [50.0, 50.0]
S2RCQL	74.5 [83.5, 85.6]	63.8 [74.6, 73.8]	63.8 [72.5, 73.4]	60.6 [71.6, 69.6]	55.6 [63.8, 64.7]	56.7 [67.4, 65.7]	50.0 [53.3, 56.7]	50.0 [62.5, 58.9]

Experimental results on robotic navigation tasks.The best results are highlighted by Bold, and the best baseline models are marked by Underline. The score in each cell is represented as "Success Rate[Optimality Rate]".

Method	ChatGPT 3.5	ERNIE-Bot 4.0	ChatGLM-6B
Naive	11.0 [4.5]	11.4 [4.0]	9.2 [3.5]
CoT	15.0 [5.7]	15.0 [5.5]	12.5 [4.5]
ToT	16.4 [5.6]	17.0 [4.5]	12.3 [4.5]
ReAct	17.7 [10.5]	16.5 [9.8]	15.0 [7.8]
Rem	46.4 [32.6]	42.1 [30.3]	37.9 [28.6]
S2RCQL	72.5 [66.7]	74.6 [63.8]	65.3 [56.7]