The Oregon State University has invented Cassie, a bipedal robot that has achieved a historic feat of running 5kms in only 53 minutes. Though it is not close to the world record set by Joshua Cheptegei,Cassie’s self-learning capability looks like it will be a formidable contender.
Cassie is built over 16 months under the guidance of Professor Jonathan Hurst , the co-founder of the Oregon State University Robotics Institute and the CTO of Agility Robotics. The Defense Advanced Research Projects Agency funded the project.
Introduced as a part of the robotic-legged mobility project in 2017, The students of the OSU explore cassie’s machine learning options.
Cassie is the first-ever bipedal robot that can sprint far distances on a single charge. This achievement marks the bipedal robot to instill all sorts of machine learning to run on flat terrain.
Experts could achieve this by striking the right combination between biomechanics and detailed insights on controlling the robots. Specific Machine Learning tools were incorporated, with Cassie’s intelligent mechanical blueprint considered ML’s ideal foundation.
ATRIAS, Digit, and Cassie
The Dynamics Robotics Laboratory first developed a robot called ATRIAS. They, later on, built another version of the robot known as Digit, which incorporated similar features to that of Cassie.
With their knowledge of dynamics, they monitored each piece of the robot and created the perfect bipedal robot that could run, jump, and do several other things. Since that robot could move and perform similar motions, the team decided to create an excellent and advanced version of ATRIAS to attract people’s attention.
Cassie can climb stairs and unfriendly terrains
We have seen several other researchers trying to train legged robots to walk like humans in challenging terrains.
About thousands of policies of continuous interaction with a flight of stairs had to be mastered before Cassie could reliably climb up and down the stairs. Once achieved, the walking robot could ascend and descend stairs without being familiar with its surrounding environment.
The robotic experts had Cassie climb a flight of stairs ten times, out of which it had successfully managed to descend the staircase ten times but could only ascend eight times.
With this, the researchers concluded that such control policies could only deal with substantial singular steps stably.
With the help of such control policies, Cassie can effectively descend steep hills and unfriendly terrains. This form of control policy primarily focuses on combining three essential elements:
a sturdy swing of the foot, a reactive type of control when the robot is in a standing position, and proprioceptive sensing of the terrain below.
Cassie’s structure and learning methods
Machine learning tools were experimented with to find the right blend between machine learning and biomechanics. Alongside a few other methods to control robots, are being implemented by the students of the OSU.
By following this approach, experts believe that developers can include performances similar to animals with robots.
Unlike the Ollie Mini-Bot, which does backflips but runs on wheels, Cassie had to strike a balance with its legs.
Cassie posses a pair of knees inspired by an ostrich which can even bend like one! The robot self-taught itself how to sprint with the help of a reinforcement learning algorithm.
Since the robot doesn’t possess a very stable posture, dynamic balancing is crucial to make it run. Simultaneously, the robot had to switch positions and execute any form of motion. Regardless, Cassie has mastered the ability to achieve a perfect balance and remain stable.
Challenges faced in the development of Cassie
The Dynamics Robotics Laboratory has a primary motive for robots with legs specifically. They wish to acquire agile locomotion to make bipedal robots more helpful and involved in real life. They aim to achieve this by developing robots that can move swiftly while tackling impacts and transferring kinetic energy.
Even though it sounds complicated, Dynamics Robotics Laboratory seeks to make bipedal robots run, hop and skip without limitations. Cassie is one of their achievements since it can climb stairs, take a sprint, and walk.
Regardless of the expertise, one possesses, such tasks are difficult to achieve primarily because of hybrid, nonlinear dynamics.
An ambiguous reward specification and contact being a few of its common causes.
To move past this, experts at the Dynamics Robotics Laboratory had to blend the initial principles of legged locomotion with different control systems.
Another set of challenges the team encountered was during Cassie’s 5 kilometers sprint, where the robot misplaced its balance and fell twice. It fell for the first time due to the computer overheating and another time when the robot took a turn at a very high speed.
Scope of bipedal robots
Cassie can assist the logistics industry with delivering packages , handling warehouse chores, and providing assistance in other domains. With robots like Cassie, people can interact and work alongside them. Having these robots exist amongst us would make our lives more convenient and enhance their quality.