notesum.ai

Published at December 3

Optimizing Latent Goal by Learning from Trajectory Preference

cs.AI

cs.LG

Released Date: December 3, 2024

Authors: Guangyu Zhao¹, Kewei Lian¹, Haowei Lin¹, Haobo Fu², Qiang Fu², Shaofei Cai¹, Zihao Wang¹, Yitao Liang¹

Aff.: ¹Peking University; ²Tencent AI Lab

Arxiv: http://arxiv.org/pdf/2412.02125v1

Refer to caption

	In Distribution							Out of Distribution
Task	GRO	GRO+	$\Delta$	STE	STE+	$\Delta$		GRO	GRO+	$\Delta$	STE	STE+	$\Delta$
wood	3.14	3.62	15.3%	3.73	3.90	4.6%		3.88	4.22	8.8%	4.22	4.29	1.7%
dirt	27.0	62.8	132.6%	16.3	36.4	123.3%		15.4	54.6	254.5%	30.4	48.0	57.9%
wool	30.4	40.8	34.2%	43.3	56.6	30.7%		34.0	41.6	22.4%	45.6	60.2	32.0%
seagrass	20.2	20.8	3.0%	4.2	21.8	419.0%		7.8	9.4	20.5%	41.4	49.0	18.4%
stonecutter	31.0	44.6	43.9%	14.1	19.0	34.8%		20.0	23.4	17.0%	36.2	48.4	33.7%
ladder	5.4	10.4	92.6%	30.9	40.2	30.1%		4.4	9.6	118.2%	29.6	41.2	39.2%
enchant	15.0	18.4	22.7%	0	0	-		19.4	21.8	12.4%	0	0	-
crafting_table	5.4	14.6	170.4%	4.0	9.6	140.0%		6.0	18.4	206.7%	2.0	6.4	220.0%
mine	4.91	6.58	34.0%	6.46	7.32	13.3%		3.9	5.38	37.9%	3.49	5.37	53.9%
chest	15.7	21.2	35.0%	3.4	4.2	23.5%		38.4	38.2	-0.5%	0.5	0.6	20.0%
hunt	31.2	39.8	27.6%	2.9	1.0	-65.5%		20.8	21.6	3.8%	1	0.2	-80.0%
combat	31.7	36.6	15.5%	0	0	-		83.4	85.6	2.6%	0	0	-
plant	2.71	3.09	14.0%	1.74	1.81	4.0%		2.85	3.11	9.1%	1.79	1.94	8.4%
pumpkin	48.3	57.8	19.7%	1.3	6.2	376.9%		16.6	25.8	55.4%	7.6	14.0	84.2%
bow	77.4	85.8	10.9%	88.9	97.8	10.0%		77.4	90.6	17.1%	65.2	88.0	35.0%
flint	1.2	7.4	516.7%	73.6	76.6	4.1%		1.2	5.8	383.3%	48.0	52.0	8.3%
obsidian	42.0	57.2	36.2%	0.4	0.7	75.0%		4.2	8.2	95.2%	0	0	-