March 4, 2019

Hello There, Mini Cheetah

Now that this is officially out I can finally put up something about what I've been doing for the past 2 years.

Photo Credit: Bayley Wang

This project has been a continuation of the Hobbyking Cheetah project, but for research rather than out of my pocket.  Here's my thesis on the actuator and robot design.  I designed and built all the all the hardware, both mechanical and electronics (with some design and fabrication help from Alex), and Jared wrote pretty much all the software and high-level control running the robot, including the Convex Model Predictive Control for locomotion.

The design principles behind the robot are very similar to Cheetah 3:  High torque electric density motors, low gear ratio, efficient transmissions to minimize friction and reflected inertia, lightweight legs with motors placed to minimize leg inertia.  A few things are different, though.  Mini Cheetah has 12 identical modular actuators with built-in motor control, gearbox, and support structure.   The electric motors are off-the-shelf and very cheap, unlike Cheetah 3's custom designed motors.  Mini Cheetah has even more range of motion than Cheetah 3 at the hip, so it's able to point it's legs completely sideways.  This means that it should be possible to make the robot land feet-first regardless of what orientation it falls in.  

I finished the hardware for the robot last May, but the publication cycle is kind of slow, so I wasn't  able to put up any info about it before.  There's going to be a paper about the robot in ICRA 2019 in Montreal, so the robot and I will both be there.


Jared and I (mostly Jared) did the backflips for our final project for Underactuated Robotics last spring.  There are some more details about how it works in the thesis, but basically we did a nonlinear trajectory optimization offline, and just played back the resulting joint torques on the motors, with a little bit of joint position control to the optimized trajectory on top.  Our simulation of the dynamics was accurate enough that it worked on the first try in the hardware.

Here are the videos from back in May.  At the time we weren't set up for backflips completely untethered and wireless.  Since then we've adjusted the flip trajectory a bit, made the landing less bouncy and more reliable, and gone fully wireless.