BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Australia/Melbourne
X-LIC-LOCATION:Australia/Melbourne
BEGIN:DAYLIGHT
TZOFFSETFROM:+1000
TZOFFSETTO:+1100
TZNAME:AEDT
DTSTART:19721003T020000
RRULE:FREQ=YEARLY;BYMONTH=4;BYDAY=1SU
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:19721003T020000
TZOFFSETFROM:+1100
TZOFFSETTO:+1000
TZNAME:AEST
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260114T163702Z
LOCATION:Meeting Room C4.11\, Level 4 (Convention Centre)
DTSTART;TZID=Australia/Melbourne:20231212T142500
DTEND;TZID=Australia/Melbourne:20231212T144000
UID:siggraphasia_SIGGRAPH Asia 2023_sess120_papers_841@linklings.com
SUMMARY:Neural Categorical Priors for Physics-Based Character Control
DESCRIPTION:Qingxu Zhu, He Zhang, Mengting Lan, and Lei Han (Tencent)\n\nR
 ecent advances in learning reusable motion priors have demonstrated their 
 effectiveness in generating naturalistic behaviors. In this paper, we prop
 ose a new learning framework in this paradigm for controlling physics-base
 d characters with significantly improved motion quality and diversity over
  existing state-of-the-art methods. The proposed method uses reinforcement
  learning (RL) to initially track and imitate life-like movements from uns
 tructured motion clips using the discrete information bottleneck, as adopt
 ed in the Vector Quantized Variational AutoEncoder (VQ-VAE). This structur
 e compresses the most relevant information from the motion clips into a co
 mpact yet informative latent space, i.e., a discrete space over vector qua
 ntized codes. By sampling codes in the space from a trained categorical pr
 ior distribution, high-quality life-like behaviors can be generated, simil
 ar to the usage of VQ-VAE in computer vision. Although this prior distribu
 tion can be trained with the supervision of the encoder's output, it follo
 ws the original motion clip distribution in the dataset and could lead to 
 imbalanced behaviors in our setting. To address the issue, we further prop
 ose a technique named prior shifting to adjust the prior distribution usin
 g curiosity-driven RL. The outcome distribution is demonstrated to offer s
 ufficient behavioral diversity and significantly facilitates upper-level p
 olicy learning for downstream tasks. We conduct comprehensive experiments 
 using humanoid characters on two challenging downstream tasks, sword-shiel
 d striking and two-player boxing game. Our results demonstrate that the pr
 oposed framework is capable of controlling the character to perform consid
 erably high-quality movements in terms of behavioral strategies, diversity
 , and realism.\n\nRegistration Category: Full Access\n\nSession Chair: Jun
 gdam Won (Seoul National University)\n\n
URL:https://asia.siggraph.org/2023/full-program?id=papers_841&sess=sess120
END:VEVENT
END:VCALENDAR