[Homepage]   [White Paper]   [Documentations]   [ARISE Initiative]

This robosuite fork contains the robosuite locomotion project for CS391 Robot Learning.

• Added Support for quadruped locomotion simulation
• Added two quadruped robots A1 and Laikago
• Added tasks, Walk, StandUp, and ClimbStairs (reward function was not defined at all for ClimbStairs, for the two other ones reward shaping requires tuning)
• For now to control the robot, a workaround controller "LOCOMOTION_JOINT_TORQUE" was added. It is a hack, and the BaseController class shoud be modified to accomodate controls for both quadruped and manipulation?

Since this is a robosuite fork, set-up the code and install just like the origial repository, please see here for details: https://robosuite.ai/docs/installation.html

In order to use the new robots, you can put the robot names "Laikago" or "A1", in the "robots" argument of the suite.make function. The robots "A1" and "Laikago" will work only with the environment "Walk","StandUp" and "ClimbStairs". The other environments are intended for armed robots. Here is a minimalistic example:

import numpy as np import robosuite as suite from robosuite import load_controller_config controller_conf = load_controller_config(default_controller="LOCOMOTION_JOINT_TORQUE") # create environment instance env = suite.make( env_name="Walk", # try with other tasks like "StandUp" or "ClimbStairs" robots="A1", # try with "Laikago" controller_configs=controller_conf, has_renderer=True, render_camera="frontview", has_offscreen_renderer=False, use_camera_obs=False, control_freq=10, init_robot_pose=[0., 0., 0.45] ) # reset the environment env.reset() for i in range(1000): action = np.random.randn(env.robots[0].dof) # sample random action obs, reward, done, info = env.step(action) # take action in the environment env.render() # render on display 

The RL part was done with the script './robosutie/test_rl/RL.py'. Nothing fancy there, but in order to train an agent with max horizon of 100 the SAC algorithm do:

python RL.py -train -alg SAC -horizon 100

To modify the environment and/or robot change the 'env_name' and 'robots' variables in the suite.make function inside RL.py.

To run a execution and visualization of the policy learned with the previous command do: python RL.py -alg SAC -horizon 100

To run an execution of some other save model do: python RL.py -alg SAC -filename <path to pickle zip file> -horizon <X>

The Reinforcement Learning was done using the stable-baselines3 library, so you'll need to install it first in order to run the script above: https://stable-baselines3.readthedocs.io/en/master/guide/install.html