notesum.ai

Published at December 10

Mobile-TeleVision: Predictive Motion Priors for Humanoid Whole-Body Control

cs.RO

cs.AI

cs.LG

Released Date: December 10, 2024

Authors: Chenhao Lu¹, Xuxin Cheng¹, Jialong Li¹, Shiqi Yang¹, Mazeyu Ji¹, Chengjing Yuan¹, Ge Yang², Sha Yi¹, Xiaolong Wang³

Aff.: ¹UC San Diego; ²MIT; ³NVIDIA

Arxiv: http://arxiv.org/pdf/2412.07773v1

[Uncaptioned image]

Method	$\mathrm{E}_{\text{acc}}^{\text{upper}}$ $\downarrow$	$\mathrm{E}_{\text{action}}^{\text{upper}}$ $\downarrow$	$\mathrm{E}_{\text{kpe}}^{\text{upper}}$ $\downarrow$	$\mathrm{E}_{\text{jpe}}^{\text{upper}}$ $\downarrow$	$\mathrm{E}_{\text{vel}}$ $\downarrow$	$\mathrm{E}_{\text{ang}}$ $\downarrow$	$\mathrm{E}_{\text{acc}}^{\text{lower}}$ $\downarrow$	$\mathrm{E}_{\text{action}}^{\text{lower}}$ $\downarrow$	$\mathrm{E}_{\text{g}}$ $\downarrow$
(a) Simulation results for H1
Ours (PMP)	5.07	0.00280	0.02161	0.02355	0.01296	0.01947	13.28	0.01651	1.301
Exbody	11.93	0.01595	0.02303	0.03125	0.00784	0.02105	14.69	0.01651	1.367
Exbody (Whole)	9.97	0.01576	0.02474	0.03788	0.00782	0.02051	14.76	0.01553	1.454
Ours w/o motion prior	5.24	0.00282	0.02146	0.02425	0.01349	0.02081	13.27	0.01623	1.585
Ours w/o curriculum	5.39	0.00289	0.02150	0.02440	0.01481	0.02057	13.54	0.01645	1.597
(b) Simulation results for GR1
Ours (PMP)	3.08	0.00054	0.02165	0.00730	0.00409	0.00886	7.83	0.00784	0.420
Exbody (Reimplementation)	7.47	0.00200	0.02282	0.01006	0.00978	0.01448	9.74	0.00425	0.835
Ours w/o motion prior	3.11	0.00054	0.02166	0.00730	0.00393	0.00899	8.07	0.00773	0.442
Ours w/o curriculum	3.38	0.00055	0.02166	0.00737	0.00424	0.01105	9.24	0.00916	0.546