Embedded Files
AI For EchoArena (UC Berkeley) aimed at recreating elementary training scenarios for the game EchoArena and presenting them in a personalized sequence to individual trainees. With a well established knowledge tracing model, AI can learn how well a trainee did at mastering certain skills, and adjust the training curriculum at runtime.
AI For EchoArena is a research project associated with UC Berkeley by Edward Kim, Zachary A. Pardos, Alton P Sturgis, Kyle Cui, James W Hu, Yunzhong Xiao, Boxi Fu, Daniel He, Isaac Gonzales, Bjoern Hartmann, Alberto Sangiovanni-Vincentelli, and Sanjit A Seshia. This is a research paper submitted to the Computer Human Interaction Conference for 2023.
Roles: Programmer, Technical Artist, Animator, Designer
As a programmer, I coded:
As a programmer, I coded:
knowledge tracing model and curriculum generation system
elementary training scenes using Scenic programming language
metrics data collection during training
As an technical artist, I created:
As an technical artist, I created:
non-player character movement and action animation
interaction with disk, environment, and player
EchoArena map and 3D environment remodeling
ABSTRACT:
Psychomotor skills, which consist of physical movements with conscious cognitive processing, are fundamental for users to engage in physically interactive extended reality (XR). These skills range from simple hand gestures to tactical movements, depending on XR applications users engage in. However, existing approaches in XR are not personalized enough to assist users to generalize their learned skills. This could constrict the XR user population. Hence, we build an intelligent, personalized XR training system that uses structured variability, or a distribution, in training tasks to induce better generalization of psychomotor skills. We formally model distributions of training tasks as a set of probabilistic programs to train a set of skills. Our system incorporates bayesian knowledge tracing (BKT) to predict when a user achieves mastery of each skill with respect its associated task distribution. Using BKT’s predictions, our system personalizes curriculum via adaptive sequencing of probabilistic programs. It further personalizes the number of tasks, or practices, to sample from each program to help each user achieve skill mastery. We conduct a user study to investigate the effectiveness of our system and BKT as a crucial design component to personalize training.
Contact and Links
Contact and Links
(240) 310 3850
Page updated
Google Sites
Report abuse