a few weeks ago I've been working on the kinematic modeling of Nybble and was deriving gaits like walking climbing etc. with this method (see the other post "Kinematic Model"). The problem was, that without further physical description, the kinematic model is to idealistic. Normally Nybble shifts its weight and there will be some kind of momentum leading to unpredictable movements. Therefore I concentrated on extending the kinematic model with a physical description in a simulation environment called pybullet.
So first, let's take a look at the result:
The model is based on an URDF File (Unified Robotics Description Format) which is a common file format that can be implemented in different simulation environments. Please take a closer look at the attached URDF-file for more information. It's mostly self explanatory. One Note: Masses, inertia, friction damping and so on are a first estimation and certainly not correct. If you have better data, please update the information in the URDF-file and upload it to the forum.
I copied the walking gait from the instincts.h into the environment and controlled the position of the joints and links to make Nybble walk. This makes it interesting to test new gaits before implementing it on Nybble and creates many more possibilities.
To make it run, you will have to install pybullet via python pip and also the bullet environment. Please see the homepage for further details: https://pybullet.org/wordpress/
Here's the code for the pybullet simulation:
And this is the necessary URDF-file:
(Please rename to nybble.urdf.I couldn't upload it with an URDF ending.)
I will provide further code updates on:
.... and even better, it would be to create a step-by-step guide, also made up of screenshoots, where you can better understand what the procedure is in order to generate a pace and apply it to our robots.
It would also be useful for the trade of the Petoi company
Ma non esiste un applicativo dove possibile , dopo aver creato un andatura, esportare il file in python ?
The gait was derived from an inverse kinematic model with a certain swing-stance pattern (see my other post kinematic model).
Pybullet has the AI gym for deep learning - there are a few examples in their github repository. I didn‘t try it yet, but I am planning to do so.
Wow super cool ! I was looking for information on this kind of simulator.
Does he learn by reinforcement ?