Imitation Learning with Petoi Bittle
Imitation Learning with Petoi Bittle
Teach Bittle new skills with simulations from real actions from dogs
What is Imitation Learning?
Quadruped robots such as the Petoi Bittle can be taught complex movements by capturing the real life movements in an animal (such as a tag), and then simulating and transferring those movements to a robot such as the Petoi Bittle. Imitation learning (or teaching a robot skills by imitating a real animal) has become common in the last 2 years and has been applied on many commercial robots, such as to a Laikago robot.
Apply Imitation Learning to Petoi Bittle
Researchers from University of Virginia and Harvard recently tried Imitation Learning on the Petoi Bittle. Their work is available here. The researchers found several limitations in using imitation learning to teach Bittle complex skills, and tried to overcome them. Here are some examples.
Low cost servos: Bittle uses servos (electric motors) in its joints to achieve precise movements. However, servos used in Bittle do not have encoders to indicate their precise joint position. To solve this limitation, researchers generated models with lower dimensions, and did not use the current joint position in their models.
Computational Limitaions: The best onboard computation that Bittle can have is the Raspberry Pi. The researchers used Tiny Machine Learning Models for which inference can be done on a Raspberry Pi.
Controllability: Bittle’s servos can be programmed for 1 degree changes. Moreover, the control unit responds to the next movement command only after the previous command is complete. To accomodate this, the researchers designed the machine learning models to yield large outputs, and smoothed out the results using a trailing average.
End Result
It is impressive that Bittle could be trained to mimic movements. The following video which shows Bittle’s performance is certainly interesting and inspiring.