OpenArms_Research_Project

OPENARMS MK.2

This product has a tiny camera in the center of the palm. The camera's purpose is that recognize object and grab it which already defined optimized movement. We received a 2nd award at 2017 Wearable Computer Contest (WCC) hosted by KAIST.

Requirements

Hardware

• Raspberry Pi 3 (Ubuntu Mate 16.04)
• Arduino nano (Arduino Uno also fine)
• Adafruit 16-Channel 12-bit PWM/Servo Driver - I2C interface - PCA9685
• Servo Motor (Gotech-SER0011 x 9ea)
• Li-Po Battery 2 cells(7.4V) 1300mAh
• spy camera for Pi
• Rotary Switch (DFRobot-SEN0156)

Software

• Opencv >= 3.2.0 (Add enable gstreamer option when you build it.)
• Tensorflow >= 1.1.0
• Keras >= 2.0.8
• ROS kinetic
• rosserial_arduino (arduino ros module)
• Adafruit-PWM-Servo-Driver-Library (https://github.com/adafruit/Adafruit-PWM-Servo-Driver-Library)

Components

Sources	Explanation
arduino/	Arduino code directory.
rospy/openarms_percepton/	detection package with model and pretrained weights.

Installation

Step 1 : Copy openarms_perception package in rospy directory to your catkin workspace src directory.

Step 2 : Upload openarms_control.ino code to your Arduino.

Please contact to design head if you need prosthetic arm design assets.

Quick Start

Step 1 : Execute roscore.

roscore

Step 2 : Launch ros module.

roslaunch openarms_perception detection.launch

After 1-2 minutes, ready message are going to print out on your screen.

Step 3 : Start ros serial communication!

rosrun rosserial_python serial_node.py /port/you/connected

Default setting of port/you/connected might be /dev/ttyUSB0.

rosrun rosserial_python serial_node.py /dev/ttyUSB0

Detection Model

Our approach is using deep learning so we chose Tiny Yolo which has lighter and faster than any other yolo model.

Detection

• We use tiny yolo trained by ms coco dataset.
• Model is for Keras framework. (backend tensorflow)
• Detection procedure takes about 2.7 seconds.

Why Tiny Yolo?

• RAM size of Raspberry Pi3 is only 1 GB.

• We need fastest detection model.

  

Weights are transformed from official site of yolo.

Model	mAP	FLOPS	keras_weights
Tiny YOLO	23.7	5.41 Bn	weights

Contact to Team LIMB

Anthony Kim : artit.anthony@gmail.com - Perception Lead

Ethan Kim : 4artit@gmail.com - Team Leader

WonJae Ji : jiwi1005@gmail.com(Head of Design)