Sensorimotor learning with the Sphero robot
05/2014 - 2017, Oswald Berthold
Python code from the Closed-loop sphero behaviour paper [1].
Dependencies
This is built on top of the following stacks, make sure to have that installed (tested only on Ubuntu)
- python-numpy, python-bluetooth, ... Use apt-get / distro mechanism
- ROS base install, use apt-get
- sphero_ros driver from [2], install via $ROSDISTRO_workspace catkin_make. See the ROS wiki how to set up a catkin workspace.
- get smp_msgs [3] and install it into your ROS workspace
- get smp_base from https://github.com/x75/smp_base, then do
export PYTHONPATH=../smp_base:$PYTHONPATH - pybluez, the python bluetooth
pip install pybluez
This is the launch sequence for the sphero ROS node
roscore
python src/sphero_ros/sphero_node/nodes/sphero.py
`sphero.py` accepts arguments such as sensorimotor loop frequency and also a BT target address to connect directly to a given device, if you know it. The experiments are tuned to 20 Hz sampling rate so we need to do
python src/sphero_ros/sphero_node/nodes/sphero.py --freq 20
or
python src/sphero_ros/sphero_node/nodes/sphero.py --freq 20 --target_addr sp:he:ro:sa:dd:re
Scripts
When roscore and sphero.py are up and running, you can run ROS clients like e.g. sphero_joystick.py
rosrun joy joy_node
A few of the utils/tests to check the basic communication is working
| Utils | |
|---|---|
| bluerssi.py | Log BT rssi values |
| sphero_colors.py | Basic color actuation with ros |
| sphero_joystick.py | Control sphero via ros joy_node |
| sphero_raw.py | Minimal test of raw communication protocol |
| sphero_simple_openloop.py | Simple open-loop command test |
| sphero_test.py | Another minimal test |
Learning
When that works, you could try the learning experiments, just to disentangle possible sources of errors.
| Experiments | |
|---|---|
| sphero_res_learner_1D.py | |
| sphero_res_learner_2D_polar.py | |
| sphero_res_learner_2D.py | |
| sphero_res_learner_1D_analyze.py | |
| sphero_data_recorder.py | |
| hk2.py | Homeostasis and homeokinesis from lpzrobots, |
| Der & Martius, 201, Playful machines |
python sphero_res_learner_1D.py --config default.cfg
You can copy the default.cfg and start editing it to play around with different targets and parameters.
Or try the homeokinesis example and play with self.*_gain parameters (in the code)
python hk2.py --mode hk --loop_time 0.05
Footnotes
[1] Berthold and Hafner, 2015, Closed-loop acquisition of behaviour on the Sphero robot, https://mitpress.mit.edu/sites/default/files/titles/content/ecal2015/ch084.html