The goal of this project is to train an open-source 3D printed quadruped robot exploring
Reinforcement Learning and OpenAI Gym. The aim is to let the robot learns domestic and generic tasks in the simulations and then
successfully transfer the knowledge (Control Policies) on the real robot without any other manual tuning.
This project is mostly inspired by the incredible works done by Boston Dynamics.
rexctl - A CLI application to bootstrap and control Rex robot running the trained Control Policies.
rex-models - A web URDF visualizer. Collection of Rex robot models.
This repository contains a collection of OpenAI Gym Environments used to train Rex, the Rex URDF model,
the learning agent implementation (PPO) and some scripts to start the training session and visualise the learned Control Polices.
This CLI application allows batch training, policy reproduction and single training rendered sessions.
Create a Python 3.7 virtual environment, e.g. using Anaconda
conda create -n rex python=3.7 anaconda
conda activate rex
Install the public rex-gym package:
pip install rex_gym
Clone this repository and run from the root of the project:
pip install .
Run rex-gym --help to display the available commands and rex-gym COMMAND_NAME --help to show the help
message for a specific command.
Use the --arg flag to eventually set the simulation arguments. For a full list check out the environments parameters.
To switch between the Open Loop and the Bezier controller (inverse kinematics) modes, just append either the --open-loop or --inverse-kinematics flags.
rex-gym COMMAND_NAME -ik
rex-gym COMMAND_NAME -ol
For more info about the modes check out the learning approach.
To start a pre-trained agent (play a learned Control Policy):
rex-gym policy --env ENV_NAME
To start a single agent rendered session (agents=1, render=True):
rex-gym train --playground True --env ENV_NAME --log-dir LOG_DIR_PATH
To start a new batch training session:
rex-gym train --env ENV_NAME --log-dir LOG_DIR_PATH
The robot used for this first version is the Spotmicro made by Deok-yeon Kim.
I've printed the components using a Creality Ender3 3D printer, with PLA and TPU+.
The hardware used is listed in this wiki.
The idea is to extend the robot adding components like a robotic arm on the top of the rack and a LiDAR sensor in the next versions alongside fixing some design issue to support a better (and easier) calibration and more reliable servo motors.
Rex is a 12 joints robot with 3 motors (Shoulder, Leg and Foot) for each leg.
The robot base model is imported in pyBullet using an URDF file.
The servo motors are modelled in the model/motor.py class.
The arm model has the open source 6DOF robotic arm Poppy Ergo Jr equipped on the top of the
rack.
To switch between base and arm models use the --mark flag.
This library uses the Proximal Policy Optimization (PPO) algorithm with a hybrid policy defined as:
a(t, o) = a(t) + π(o)
It can be varied continuously from fully user-specified to entirely learned from scratch.
If we want to use a user-specified policy, we can set both the lower and the upper bounds of π(o) to be zero.
If we want a policy that is learned from scratch, we can set a(t) = 0 and give the feedback component π(o) a wide output range.
By varying the open loop signal and the output bound of the feedback component, we can decide how much user control is applied to the system.
A twofold approach is used to implement the Rex Gym Environments: Bezier controller and Open Loop.
The Bezier controller implements a fully user-specified policy. The controller uses the Inverse Kinematics model (see model/kinematics.py)
to generate the gait.
The Open Loop mode consists, in some cases, in let the system lean from scratch (setting the open loop component a(t) = 0) while others
just providing a simple trajectory reference (e.g. a(t) = sin(t)).
The purpose is to compare the learned policies and scores using those two different approach.
This is the list of tasks this experiment want to cover:
- Basic controls:
- Static poses - Frame a point standing on the spot.
- Bezier controller
- Open Loop signal
- Gallop
- forward
- Bezier controller
- Open Loop signal
- backward
- Bezier controller
- Open Loop signal
- Walk
- forward
- Bezier controller
- Open Loop signal
- backward
- Bezier controller
- Open Loop signal
- Turn - on the spot
- Bezier controller
- Open Loop signal
- Stand up - from the floor
- Bezier controller
- Open Loop signal
- Navigate uneven terrains:
- Random heightfield, hill, mount
- Maze
- Stairs
- Open a door
- Grab an object
- Fall recovery
- Reach a specific point in a map
- Map an open space
To set a specific terrain, use the --terrain flag. The default terrain is the standard plane. This feature is quite useful to
test the policy robustness.
Use the --terrain random flag to generate a random heighfield pattern. This pattern is updated at every 'Reset' step.
Use the --terrain hills flag to generate an uneven terrain.
Use the --terrain mounts flag to generate this scenario.
Use the --terrain maze flag to generate this scenario.
Goal: Move Rex base to assume static poses standing on the spot.
The gym environment is used to learn how to gracefully assume a pose avoiding too fast transactions.
It uses a one-dimensional action space with a feedback component π(o) with bounds [-0.1, 0.1].
The feedback is applied to a sigmoid function to orchestrate the movement.
When the --playground flag is used, it's possible to use the pyBullet UI to manually set a specific pose altering the robot base position
(x,y,z) and orientation (roll, pitch, jaw).
Goal: Gallop straight on and stop at a desired position.
In order to make the learning more robust, the Rex target position is randomly chosen at every 'Reset' step.
This gym environment is used to learn how to gracefully start the gait and then stop it after reaching the target position (on the x axis).
It uses two-dimensional action space with a feedback component π(o) with bounds [-0.3, 0.3]. The feedback component is applied to two ramp functions
used to orchestrate the gait. A correct start contributes to void the drift effect generated by the gait in the resulted learned policy.
This gym environment is used to let the system learn the gait from scratch. The action space has 4 dimensions, two for the front legs and feet
and two for the rear legs and feet, with the feedback component output bounds [−0.3, 0.3].
Goal: Walk straight on and stop at a desired position.
In order to make the learning more robust, the Rex target position is randomly chosen at every 'Reset' step.
This gym environment is used to learn how to gracefully start the gait and then stop it after reaching the target position (on the x axis).
It uses two-dimensional action space with a feedback component π(o) with bounds [-0.4, 0.4]. The feedback component is applied to two ramp functions
used to orchestrate the gait. A correct start contributes to void the drift effect generated by the gait in the resulted learned policy.
This gym environment uses a sinusoidal trajectory reference to alternate the Rex legs during the gait.
leg(t) = 0.1 cos(2π/T*t)
foot(t) = 0.2 cos(2π/T*t)
The feedback component has very small bounds: [-0.01, 0.01]. A ramp function are used to start and stop the gait gracefully.
Goal: Reach a target orientation turning on the spot.
In order to make the learning more robust, the Rex start orientation and target are randomly chosen at every 'Reset' step.
This gym environment is used to optimise the step_length and step_rotation arguments used by the GaitPlanner to implement the 'steer' gait.
It uses a two-dimensional action space with a feedback component π(o) with bounds [-0.05, 0.05].
This environment is used to learn a 'steer-on-the-spot' gait, allowing Rex to moving towards a specific orientation.
It uses a two-dimensional action space with a small feedback component π(o) with bounds [-0.05, 0.05] to optimise the shoulder and foot angles
during the gait.
Goal: Stand up starting from the standby position
This environment introduces the rest_postion, ideally the position assumed when Rex is in standby.
The action space is equals to 1 with a feedback component π(o) with bounds [-0.1, 0.1] used to optimise the signal timing.
The signal function applies a 'brake' forcing Rex to assume an halfway position before completing the movement.
| Environment | env flag |
arg flag |
|---|---|---|
| Galloping | gallop | target_position |
| Walking | walk | target_position |
| Turn | turn | init_orient, target_orient |
| Stand up | standup | N.A |
arg |
Description |
|---|---|
| init_orient | The starting orientation in rad. |
| target_orient | The target orientation in rad. |
| target_position | The target position (x axis). |
| Flags | Description |
|---|---|
| log-dir | The path where the log directory will be created. (Required) |
| playground | A boolean to start a single training rendered session |
| agents-number | Set the number of parallel agents |
You may want to edit the PPO agent's default configuration, especially the number of parallel agents launched during the simulation.
Use the --agents-number flag, e.g. --agents-number 10.
This configuration will launch 10 agents (threads) in parallel to train your model.
The default value is setup in the agents/scripts/configs.py script:
def default():
"""Default configuration for PPO."""
# General
...
num_agents = 20
Sim-to-Real: Learning Agile Locomotion For Quadruped Robots and all the related papers. Google Brain, Google X, Google DeepMind - Minitaur Ghost Robotics.
Inverse Kinematic Analysis Of A Quadruped Robot
Leg Trajectory Planning for Quadruped Robots with High-Speed Trot Gait
Deok-yeon Kim creator of SpotMini.
The awesome Poppy Project.
SpotMicro CAD files: SpotMicroAI community.
The kinematics model was inspired by the great work done by Miguel Ayuso.
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent