Deep RL Arm Manipulation

This project is based on the Nvidia open source project "jetson-reinforcement" developed by Dustin Franklin. The goal of the project is to create a DQN agent and define reward functions to teach a robotic arm to carry out two primary objectives:

1. Have any part of the robot arm touch the object of interest, with at least a 90% accuracy.
2. Have only the gripper base of the robot arm touch the object, with at least a 80% accuracy.

Building from Source (Nvidia Jetson TX2)

Run the following commands from terminal to build the project from source:

$ sudo apt-get install cmake
$ git clone http://github.com/udacity/RoboND-DeepRL-Project
$ cd RoboND-DeepRL-Project
$ git submodule update --init
$ mkdir build
$ cd build
$ cmake ../
$ make

During the cmake step, Torch will be installed so it can take awhile. It will download packages and ask you for your sudo password during the install.

Testing the API

To make sure that the reinforcement learners are still functioning properly from C++, a simple example of using the API called catch is provided. Similar in concept to pong, a ball drops from the top of the screen which the agent must catch before the ball reaches the bottom of the screen, by moving it's paddle left or right.

To test the textual catch sample, run the following executable from the terminal. After around 100 episodes or so, the agent should start winning the episodes nearly 100% of the time:

$ cd RoboND-DeepRL-Project/build/aarch64/bin
$ ./catch
[deepRL]  input_width:    64
[deepRL]  input_height:   64
[deepRL]  input_channels: 1
[deepRL]  num_actions:    3
[deepRL]  optimizer:      RMSprop
[deepRL]  learning rate:  0.01
[deepRL]  replay_memory:  10000
[deepRL]  batch_size:     32
[deepRL]  gamma:          0.9
[deepRL]  epsilon_start:  0.9
[deepRL]  epsilon_end:    0.05
[deepRL]  epsilon_decay:  200.0
[deepRL]  allow_random:   1
[deepRL]  debug_mode:     0
[deepRL]  creating DQN model instance
[deepRL]  DQN model instance created
[deepRL]  DQN script done init
[cuda]  cudaAllocMapped 16384 bytes, CPU 0x1020a800000 GPU 0x1020a800000
[deepRL]  pyTorch THCState  0x0318D490
[deepRL]  nn.Conv2d() output size = 800
WON! episode 1
001 for 001  (1.0000)
WON! episode 5
004 for 005  (0.8000)
WON! episode 10
007 for 010  (0.7000)
WON! episode 15
010 for 015  (0.6667)
WON! episode 20
013 for 020  (0.6500)  13 of last 20  (0.65)  (max=0.65)
WON! episode 25
015 for 025  (0.6000)  11 of last 20  (0.55)  (max=0.65)
LOST episode 30
018 for 030  (0.6000)  11 of last 20  (0.55)  (max=0.65)
LOST episode 35
019 for 035  (0.5429)  09 of last 20  (0.45)  (max=0.65)
WON! episode 40
022 for 040  (0.5500)  09 of last 20  (0.45)  (max=0.65)
LOST episode 45
024 for 045  (0.5333)  09 of last 20  (0.45)  (max=0.65)
WON! episode 50
027 for 050  (0.5400)  09 of last 20  (0.45)  (max=0.65)
WON! episode 55
031 for 055  (0.5636)  12 of last 20  (0.60)  (max=0.65)
LOST episode 60
034 for 060  (0.5667)  12 of last 20  (0.60)  (max=0.65)
WON! episode 65
038 for 065  (0.5846)  14 of last 20  (0.70)  (max=0.70)
WON! episode 70
042 for 070  (0.6000)  15 of last 20  (0.75)  (max=0.75)
LOST episode 75
045 for 075  (0.6000)  14 of last 20  (0.70)  (max=0.75)
WON! episode 80
050 for 080  (0.6250)  16 of last 20  (0.80)  (max=0.80)
WON! episode 85
055 for 085  (0.6471)  17 of last 20  (0.85)  (max=0.85)
WON! episode 90
059 for 090  (0.6556)  17 of last 20  (0.85)  (max=0.85)
WON! episode 95
063 for 095  (0.6632)  18 of last 20  (0.90)  (max=0.90)
WON! episode 100
068 for 100  (0.6800)  18 of last 20  (0.90)  (max=0.90)
WON! episode 105
073 for 105  (0.6952)  18 of last 20  (0.90)  (max=0.90)
WON! episode 110
078 for 110  (0.7091)  19 of last 20  (0.95)  (max=0.95)
WON! episode 111
079 for 111  (0.7117)  19 of last 20  (0.95)  (max=0.95)
WON! episode 112
080 for 112  (0.7143)  20 of last 20  (1.00)  (max=1.00)

Internally, catch is using the dqnAgent API from our C++ library to implement the learning.

Project Environment

To get started with the project environment, run the following:

$ cd RoboND-DeepRL-Project/build/aarch64/bin
$ chmod u+x gazebo-arm.sh
$ ./gazebo-arm.sh

The plugins which hook the learning into the simulation are located in the gazebo/ directory of the repo. The RL agent and the reward functions are to be defined in ArmPlugin.cpp.