notesum.ai

Published at December 4

Diffusion-VLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression

cs.RO

cs.CV

Released Date: December 4, 2024

Authors: Junjie Wen¹, Minjie Zhu², Yichen Zhu³, Zhibin Tang, Jinming Li, Zhongyi Zhou, Chengmeng Li, Xiaoyu Liu, Yaxin Peng, Chaomin Shen, Feifei Feng

Aff.: ¹East China Normal University; ²Midea Group; ³Shanghai University

Arxiv: http://arxiv.org/pdf/2412.03293v1

	Pre-trained	In-Distribution						Visual Generalization
Model $\setminus$ Tasks	Trajectory	Task 1	Task 2	Task 3	Task 4	Task 5	Avg.	Task 1	Task 2	Task 3	Task 4	Task 5	Avg.
Diffusion Policy [10]	-	66.7	36.4	0	36.4	0	27.9	11.1	11.1	0	22.2	0	8.9
ScalingDP-1B [74]	-	66.7	36.4	27.3	37.0	18.2	35.2	33.3	33.3	0	22.2	22.2	22.2
TinyVLA [57]	-	72.7	45.5	36.4	45.5	27.3	45.5	44.4	44.4	11.1	22.2	22.2	28.9
Octo [34]	970K	57.6	27.3	9.1	0	27.3	24.3	44.4	11.1	11.1	0	22.2	17.8
OpenVLA-7B [26]	970K	69.7	18.2	18.2	36.4	54.5	39.4	55.6	11.1	0	33.3	33.3	26.7
DiVLA-2B	39K	100	100	63.6	63.6	90.9	83.6	44.4	66.7	44.4	66.7	66.7	57.8
DiVLA-7B	39K	100	100	72.7	81.8	100	90.9	66.7	88.9	55.6	77.8	66.7	71.0