1 of 8

Vertical Farming Automation�Progress Review 7

Team G

1

Date: Sep 7th 2022

2 of 8

Progress Review 7 Goals

​

2

1. Complete new test environment setup

​

2. Complete new tool design

​

3. Migrate from using hard-coded offsets for arm translation to using the new Tool-Integrated URDF.

​

4. Identify problems with the Hello robot and contact Hello Team for potential solutions.

​

5. Split planning and control in the navigation subsystem

​

6. Demonstrate Yolo running on Jetson Xavier.

3 of 8

Test Environment Setup

Previous Status:

  • Real plants withered over the summer
  • Unreliable availability of tomatoes on demonstration days
  • Upkeep in itself a project

Current Status:

  • New plant stems made from floral wire
  • Foliage created using fake vine leaves to recreate same noisy backdrop
  • Reduces need for upkeep

3

4 of 8

Tool Design

  • Two alternatives being evaluated - To continue with both cutter and gripper, or to simply stick with gripper alone.
  • Trade study to be conducted with magnets for tomato attachment to stem or using celery sticks to demonstrate cutting.
  • We have shown the viability of cutting in last semester’s SVD and we’d like to iron out the kinks in our main pipe-line instead of devoting energy into showing cutting. Magnets would allow easy and repeatable testing of our system refinement.
  • First version of ‘gripper only’ tool 3D printed.

4

5 of 8

Tool-integrated URDF

Previous status:

  • Arm translation using offsets and the transform from the last link on the arm

Current status:

  • Integrated the harvester frame in Tf tree
  • Testing is still tricky because of issues related to arm translation

5

6 of 8

Hello Robot Debugging

  • Hello robot charger and wacc
  • Arm extension problems
  • Async movements
  • Motor dropping commands

6

7 of 8

Navigation

Previous Status:

  • The basic pipeline has been laid but the tracking motions are very jerky.
  • Software architecture can do with significant improvements to readability and overall performance optimization.

Current Status:

  • Refactoring the existing software implementation of the navigation stack.

7

8 of 8

Perception

Previous Status:

  • Using Faster RCNN - Performance: 1~2 FPS, Detection mAP=0.53
  • Some false positives

​

Current Status:

  • Transferring from Faster RCNN to YOLO V5

→ Detection mAP=0.59�→ Expecting improved inference time on Jetson

  • Planning to optimize the model further using TensorRT�→ Installed the latest JetPack on NVIDIA’s Jetson Xavier

​

Future work:

  • Run tensorrt inference on Jetson using the trained YOLO model

8