Introducing 

Prezi AI.

Your new presentation assistant.

Refine, enhance, and tailor your content, source relevant images, and edit visuals quicker than ever before.

Loading content…
Loading…
Transcript

Adaptive Machine Learning Approaches

for Trajectory Planning

Using Genetic Algorithm and Reinforcement Learning

Enrico Marchesini VR409577

Davide Corsi VR409578

Corso di Robotica 2017/2018

Learning to Move

Machine Learning approaches for online trajectory planning

Dense reward

Success rate (x100)

Target joint angle

Target distance

Generation (x100)

Generation

Distsance (m)

Generation (x100)

Simulated using a 4DOF robotic arm

Sparse reward

Success rate (x100)

Some other issues

Generation

Robot controllers based on neural networks have the capability of performing complex tasks starting from minimal information about the structure of the robotic platform and the operational environment.

Our project uses some machine learning approaches to generate the trajectory of the Franka robotic arm.

Size problem

Simulated using a 7DOF robotic arm

4x150

2x50

Reward average

  • What timeout should we set?
  • How long the training should go?
  • Which optimizer should we use?
  • Which activation function performs better?
  • Does this network really learn something?
  • Does this network really works on the target robot arm?

Generation

2x150

Reward average

Future work

Generation

Simulated using a 4DOF robotic arm

  • Improve network performances implementing optimization algorithms on the trained model.
  • Try different input layer shaping to learn more difficult tasks with the same network.
  • Introduce a safety function for real-time obstacle avoidance.
  • Introduce a fault tolerant design to manage a joint failure online.

Reward problem

Neural Network

Genetic Algorithm

Training environment

{

Sparse:

0 on working

-1 on timeout

1 on success

Dense

Target distance

Variable target

Fixed target

Target joint angle

(Sum of joint discard)

The collaborative lightweight robot system designed specifically to assist humans.

Results

Activation function

Reinforcement Learning

Success Rate (x10)

Generation

After 18 hours training

Deep Q-Learning

Enrico Marchesini VR409577

Davide Corsi VR409578

- distance

(e -1)

Learn more about creating dynamic, engaging presentations with Prezi