MetaUrban: Rebuttal Video Demonstrations

Code | Project Page

This page displays video demonstrations in response to reviewers’ feedback. Click on any video to play and enlarge it to full screen for a better visualization. You can also find specific responses by searching the reviewer's name.

Video Index_0000
(To Reviewer FiPy and R6rL)

Traffic light rules.

Video Index_0001
(To Reviewer FiPy)

Simulation of a delivery bot moving on different terrains.

Video Index_0002
(To Reviewer FiPy)

Simulation of a delivery bot moving on different materials.

Video Index_0003
(To Reviewer FiPy)

Simulation of lane merging and diverging scenarios.

Video Index_0004
(To Reviewer R6rL)

The enhanced version of COCO robot (with stronger computation and perceiving ability) in our lab. We are collaborating with COCO Robotics for real-world food delivery.

Video Index_0005
(To Reviewer R6rL and qkEE)

Integration of Nvidia Omniverse as the renderer to improve visual realism, and Nvidia PhysX as the physical engine to improve interactive realism.

Video Index_0006
(To Reviewer R6rL and yJjR)

Preliminary results of harnessing diffusion models to improve the visual quality of MetaUrban in 2D space. Input: RGB image rendered by MetaUrban; output: photo-realistic image. (It is an extension of our previous work SimGen)

Video Index_0007
(To Reviewer R6rL and yJjR)

Preliminary results of harnessing Gaussian splatting to improve the visual quality of MetaUrban in 3D space. Input: monocular videos; output: 3D scene represented by Gaussian Splatting. Integrated within the simulator, it enables training agents with photo-realistic RGB images as observations.

Video Index_0008
(To Reviewer yJjR and qkEE)

Support of multi-agents planning task. Agents need to make joint path planning, completing point-to-point navigation with no collisions and deadlocks.

Video Index_0009
(To Reviewer yJjR and qkEE)

Support of human-in-the-loop learning task. Human can select chance to take over the control of robots when policy does not work well. This data can be used for training of human-AI shared control algorithms.