Simulating Real-World Human Activities with
VirtualCity: A Large-Scale Embodied Environment
for 2D, 3D, and Language-Driven Tasks
Name
ren-jordanr1-meng-eecs-2023-thesis.pdf
Description
Thesis PDF
Size
85.96 MB
Format
Adobe PDF
Checksum (MD5)
01ee4aac866bbbababe785592232a31f
Author(s)
Ren, Jordan
Advisor(s)
Torralba, Antonio
Date Issued
June 2023
Publisher
Massachusetts Institute of Technology
Abstract
Embodied environments act as a tool that enables various control tasks to be learned. Within these simulators, having realistic rendering and physics ensures that the sim2real gap for tasks isn’t too large. Current embodied environments focus mainly on small-scale or low-level tasks, without the capability to learn large-scale diverse tasks, and often lack the realism for a small sim2real gap. To address the shortcomings of current simulators, we propose VirtualCity, a large-scale embodied environment that enables the learning of high-level planning tasks with photo-realistic rendering and realistic physics. To interact with VirtualCity, we provide a user-friendly Python API that allows the modification, control, and observation of the environment and its agents within. Building this realistic environment brings us closer to adapting models trained in simulation to solve real-world tasks.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
