notesum.ai

Published at November 2

Prompt Tuning with Diffusion for Few-Shot Pre-trained Policy Generalization

cs.LG

cs.AI

Released Date: November 2, 2024

Authors: Shengchao Hu¹, Wanru Zhao², Weixiong Lin¹, Li Shen³, Ya Zhang¹, Dacheng Tao⁴

Aff.: ¹Shanghai Jiao Tong University; ²University of Cambridge; ³Sun Yat-sen University; ⁴Nanyang Technological University

Arxiv: http://arxiv.org/abs/2411.01168v1

Refer to caption

	Prompt-DT-Random	Prompt-DT-Expert	MT-ORL	Soft Prompt	Adaptor	Prompt-DT-FT	Prompt-Tuning DT	Prompt Diffuser	Prompt-DT-FT-Full
Sizes	-	-	-	0.24K	13.49K	13.71M	0.24K	0.17M	13.71M
Percentage	-	-	-	0.0018%	0.098%	100%	0.0018%	1.24%	100%
Cheetah-dir	935.3 $\pm$ 2.6	934.6 $\pm$ 4.4	-46.2 $\pm$ 3.4	940.2 $\pm$ 1.0	875.2 $\pm$ 4.2	936.9 $\pm$ 4.8	941.5 $\pm$ 3.2	945.3 $\pm$ 7.2	950.6 $\pm$ 1.2
Cheetah-vel	-127.7 $\pm$ 9.9	-43.5 $\pm$ 3.4	-146.6 $\pm$ 2.1	-41.8 $\pm$ 2.1	-63.8 $\pm$ 6.3	-40.1 $\pm$ 3.8	-39.5 $\pm$ 3.7	-35.3 $\pm$ 2.4	-23.0 $\pm$ 1.1
Ant-dir	278.7 $\pm$ 38.7	420.2 $\pm$ 5.1	110.5 $\pm$ 2.2	379.1 $\pm$ 33.7	361.4 $\pm$ 5.6	425.2 $\pm$ 8.6	427.9 $\pm$ 4.3	432.1 $\pm$ 6.7	494.2 $\pm$ 2.3
MW reach-v2	457.2 $\pm$ 30.3	469.8 $\pm$ 29.9	264.1 $\pm$ 9.6	448.7 $\pm$ 41.3	477.9 $\pm$ 2.1	478.1 $\pm$ 27.8	472.5 $\pm$ 29.0	555.7 $\pm$ 6.8	564.7 $\pm$ 0.7
Average	385.9	445.3	45.5	431.6	412.7	450.0	450.3	474.4	496.6