Agility Robotics announced at the Robotics Summit & Showcase two free simulators for its Cassie bipedal robot. The simulators will help evaluate the physical capabilities of the Cassie bipedal robot and test control methods.
In the opening keynote of the Robotics Summit, Agility Robotics co-founder and CEO Damion Shelton said a simulation library using MuJoCo is now available, while a simulator using Gazebo will be released in July 2018. The Gazebo release was slightly delayed to “get it right” and make it more robust.
The simulators are continuously validated against the real Cassie bipedal robot and are used by Oregon-based Agility to develop its own controllers. Shelton said Cassie is designed to behave the same in both the real and virtual worlds.
Here’s a video that shows a side-by-side comparison of outdoor walking in the real world versus simulation.
Shelton said the simulators also enable testing control methods that Agility doesn’t support in-house for the Cassie bipedal robot, such as deep learning. Testing deep learning methods on the Cassie bipedal robot in the real world would require considerable time and risks damaging the robot.
Again, because the simulators closely resemble the actual physics of the Cassie bipedal robot, deep learning testing can be performed in simulation first and later deployed to real hardware.
Here’s a video that shows how the University of British Columbia and Oregon State University used deep reinforcement learning (DRL), in simulation, for feedback control of the Cassie bipedal robot.
You can read much more about the aforementioned researchers’ work with deep reinforcement learning and the Cassie bipedal robot in this paper called “Feedback Control For Cassie With Deep Reinforcement Learning.” Here’s the premise of the paper:
“By formulating a feedback control problem as searching for an optimal imitation policy for a Markov Decision Process, we can apply DRL to train controllers for bipedal walking tasks in a model-free manner with a single reference motion. Without needing to make the model-based simplifications commonly used to tractably realize control policies, DRL is able to exploit the full dynamics of the robot and produce robust controllers. We test the robustness of our controllers by introducing sensory delays, testing blind walks on various types of terrain, and via random pushes applied to the body. Policies that can make the robot walk at a different speed can be constructed by retraining on a modified version of the reference trajectory that has been scaled in time. We can further make the robot speed up and slow down by interpolating between these policies. These results provide a degree of confidence that we can deploy a controller trained using DRL on a real biped.”
Agility raised $8 million in Series A funding in March 2018 to continue development of the Cassie bipedal robot. Playground Global led the round, with participation from Sony Innovation Fund and existing investor Robotics Hub. Bruce Leak, a founder of Playground Global, joined Agility’s board. Agility has raised a total of $8.792 million in funding.
— Steve Crowe (@SteveCrowe) May 23, 2018