search

ROBOTIS-GIT / turtlebot3_machine_learning

Apache License 2.0
123 stars 82 forks source link

Training with more laser scans #31

Closed FrancescoTerrosi closed 4 years ago

FrancescoTerrosi commented 5 years ago

Hello there, I forked this repo because I need an already-implemented neural network to implement some safety checks on it, but this is another story. Since my main focus is safety, I need the robot to get 'realistic' scan data and for that reason I expanded the state_space from 26 to 362 as pointed in #2 I am training the robot in a different environment from the ones provided by the example (i.e. turtlebot_house_world) and I noticed that, when using more laser data, the robot just keep cycling like it is about to explode, while if I don't modify the state_space (as is the default implementation) I get decent performances.

My question is: is there something in the code I need to modify in order to adapt the robot to work with more sensor data or probably it's just a matter of more training?

Thank you very much, Cheers

ROBOTIS-Will commented 5 years ago

@FrancescoTerrosi Hi, When you increase the resolution of the scan data, this could cause trouble in learning due to the expanded state_size. In order to resolve this issue, you can increase the scan data resolution little by little to find a proper resolution that works, or perform hyper parameter tuning by adjusting learning_rate, step, batch size and so on. Thank you.