sk8o-sim 0.1.4

A set of two simulation environments for robot SK8O developed at FEE CTU in Prague.

This folder contains code that to simulates SK8O in Python in two ways:

• as a linear approximation of a Segway ~ SK8O with fixed legs in sk8o_sim/segway,
• as a non-linear simulation of SK8O using MuJoCo ("full model") in sk8o_sim/full.

You can install this package from PyPI by running

pip install sk8o_sim

For more information, check my diploma thesis.

Making changes

This project targets Python 3.11, though it was originally written for Python 3.9, which may affect some parts of the codebase (particularly the typing annotations). The necessary packages can be installed using the standard requirements.txt file e.g. by

pip install -r sk8o_sim/requirements.txt

If you make any changes you wish to make available via pip, make changes to pyproject.toml as necessary (at least increment version number!) and GitLab's CI should detect this change and do the dirty work for you. :-)

Simulation

Segway

Files in sk8o_sim/segway contain the linear approximation of a Segway that was used to train RL models and to tune the LQR. It is highly configurable, can generate randomized models, simulate motor quantization and can be rendered via pygame. Note that by default, the measurement and process noise can be quite severe (in addition to initial conditions not necessarily too close to the equilibrium), so even the LQR that works just fine on SK8O may sometimes fail to stabilize the robot.

What it does not support is any way to end the simulation due to the robot falling - it is up to you to decide when the approximation is no longer valid!

Main components

The building block of the Segway simulation is the SegwaySimData dataclass found in sk8o_sim/segway/data.py. It contains all the simulation data and some helper functions. In sk8o_sim/segway/simulation.py, you can find how the simulation is performed and view.py is in charge of the rendering.

Demo

A very crude simulation of an LQR stabilization with "slow-mo" rendering can be run using the following code:

import time

import numpy as np
from sk8o_sim import (
    SegwayLQRController,
    SegwayLQRControllerCfg,
    SegwaySimulation,
    SegwaySimulationCfg,
)

# initialize everything
sim = SegwaySimulation(SegwaySimulationCfg())
data = sim.reset()
controller = SegwayLQRController(SegwayLQRControllerCfg())

# run the simulation
sim_time = 10  # [s]
for k in range(sim_time * controller.control_frequency):
    action = controller(data)
    data = sim.run(action, 1 / controller.control_frequency)
    sim.render()
    time.sleep(0.5)

Note that the FPS will likely be unstable in this example. For a more advanced interaction, check out the examples/interactive_segway.py file.

Full model

The full model is implemented in MuJoCo and can be found in sk8o_sim/full. The mujoco-python-viewer package is used for rendering.

If you've familiarized yourself with the Segway simulation, this one should hopefully feel similar. Again, you will mostly work with what's contained in the simulation.py -- this time the class is FullSimulation and will generate FullSimulationData objects found in data.py. This class is essentially a wrapper around MuJoCo that enables rendering and additionally computes the equivalent Segway states for easier control. Note that these may not make much sense whenever the robot is not in a "close-to-upright" position.

One (admittedly experimental) trick of this simulation is the option to "lock" SK8O's hips/legs (see the (un)lock_hips methods) for testing.

You may also wish to use the has_fallen method to gauge whether your control system is successful. This method will return True whenever anything else but the wheels touch the ground.

Controllers

The controllers can be found in sk8o_sim/controllers.py and include SegwayLQRController to control the wheels as per Adam's thesis and a simple PID hip controller, SK8OHipController. Both implement the SK8OController interface and are combined in the SK8OFullController class which controls both. You can use this class if you wish to implement only partial control. For example, if your controller only takes care of wheel torques, you can implement the interface and pass your controller as the wheel_controller to SK8OFullController.

Configuration

The project is ready for OmegaConf/hydra: the sk8o_sim/configs.py file contains all the configuration classes that are available for the simulation, rendering and controllers. This allows you to control everything via custom YAML configs and/or command-line arguments with a only little bit of work. This should also make logging the configuration easy. If you experiment with many versions of a new controller, you can leverage this to easily compare the changes you've made using e.g. Weights & Biases or similar software. To see a sample, see the examples/cli_configuration.py demo file.