Simulate and design “xenobots”, on the AMD platform

Submission Number: 119
Submission ID: 211
Submission UUID: 524c1673-19ff-46e8-8d6c-44ae9dedbf0c
Submission URI: /form/project

Created: Tue, 09/21/2021 - 16:50
Completed: Tue, 09/21/2021 - 17:02
Changed: Thu, 05/05/2022 - 04:18

Remote IP address: 199.168.200.4
Submitted by: Andrea Elledge
Language: English

Is draft: No
Webform: Project
Simulate and design “xenobots”, on the AMD platform
Northeast
xenobot.jpeg
administering-hpc (438), amber (292), big-data (4), biology (515), file-transfer (252), github (490), slurm (71)
Complete

Project Leader

Josh Bongard
{Empty}
{Empty}

Project Personnel

Keri Toksu
David Matthews
{Empty}

Project Information

In a previous project, undergraduate David Matthews ported our voxel-based soft-bodied physical simulator, Voxcraft, to the AMD platform. This allows us to simulate and design “xenobots” -- mm-sized robots made exclusively from genetically unmodified frog tissue —on the AMD platform.

In the new project we propose, Matthews will draw on his experience with AMD from the previous platform to port a next-generation soft-body simulator to the AMD platform. This new simulator is based on the material-point method (MPM), which is rapidly becoming the standard approach for simulating machines and robots with soft and/or biological components. This is because MPM allows state-of-art AI methods to design soft-bodied machines much more efficiently. More specifically, it allows for the backpropagation of behavioral error — where the machine “went wrong” in trying to perform the desired task — into the design and control of the machine itself. This allows the AI to avoid computationally inefficient trial and error design, and instead assume gradient-based design. We propose three deliverables for this project.

Project Information Subsection

1. In the first month, we will port the Material Point Method to the AMD cluster. This will result in the creation of a series of videos showing hand-built robots, in silico, as a connected series of points, behaving in a virtual world.

2. In the second month we will wrap a gradient-based AI method around this MPM simulator, and demonstrate how it can propagate behavioral errors back through the simulator to fix errors in the shape and control of the robot.

3. In the final month we will show how the AI method can efficiently design these robots to perform a series of simple tasks, and demonstrate that some of these AI-designed robots could actually be fabricated, in reality, as xenobots built from frog tissues.
{Empty}
{Empty}
{Empty}
Can work with any level
{Empty}
University of Vermont
89 S. Prospect Street
Burlington, Vermont. 05405
NE-University of Vermont
10/01/2021
No
Already behind3Start date is flexible
3
{Empty}
{Empty}
{Empty}
{Empty}
  • Milestone Title: Videos of hand built robots
    Milestone Description: Port the Material Point Method to the AMD cluster. This will result in the creation of a series of videos showing hand-built robots, in silico, as a connected series of points, behaving in a virtual world.
    Completion Date Goal: 2021-10-31
{Empty}
Will add questions to the portal
yes at least on paper
{Empty}
{Empty}
{Empty}
VACC cluster access to both DeepGreen and AMD gpu clusters
{Empty}

Final Report

{Empty}
{Empty}
{Empty}
From student David Matthews, BS Computer Science UVM 2021
Since the NE Cyberteam grant I have been continuing to work with professor Josh Bongard on developing a new method to efficiently optimize the topology of soft robots. We are getting close to publishing this work.
In addition I am planning to apply to Ph.D programs in AI and robotics this fall.
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}