You’re not the intended audience. This is meant for small research labs that are just starting up and want to enter the RL/Robotics research space and can’t afford a $400,000 PR2 or $150,000 shadow hand.
The main cost is the actuators. The servos they use are the only ones out there with integrated feed back from the sensors which is required so they can calculate hardware safety. They also have, albeit pretty bad, torque control which is necessary for many approaches to controlling legged robots and manipulators.
I'm skeptical that a stepper motor would be able to support it's own weight and more without a gearbox. This is especially important for making serial manipulators. Also keep in mind that what they're aiming for here is repeatability so that researchers can easily compare their algorithms. That means using parts that don't vary too much and minimizing the amount of assembly researchers have to do. There are much cheaper knock off versions of these actuators, although they may not be as repeatable. At the very least they aren't as well documented.
This is more about repeatable experiments than particular movements.
Regarding the price, the actuators seem to be ~240USD. A good stepper motor with the appropriate feedback mechanism to make it suitable for servo-like control plus a modern stepper motor controller that is suited for the robotics context will likely not be (much) cheaper and you have to hack together the servo functionality, tune settings, etc. - which seems detrimental if the goal is repeatability accross teams. I'm not in the target audience for these robots either but from the perspective of robustness and repeatable research they don't look too shabby.
Maybe I am missing something but fundamentally you can "learn" in the ML sense with any sensor and effector.
Examples of cheap sensors: MEMS microphone, camera, voltage sensor, mass sensor, distance sensor, multi-axis position sensors (gyroscope/magnetometers).
Examples of effectors: Any kind of motor, solenoid, LEDs, etc.
If you want to constrain the question to 3D motion, here is a suggestion - hack the controller of any existing RC car platform. Add overhead position sensing within a fixed arena added via external camera. Maybe add a MEMS microphone (USD$3) or position sensor (~USD$10) to verify airtime/orientation. ML problem #1: Add a ramp. Try to get it to jump highest (longest time airborne). ML problem #2: Same with a power efficiency metric. ML problem #3: Same with a time efficiency metric applied to navigation from a random start point and orientation. ML problem #4: Motor noise vs. jump height optimization.
Byron Boot's lab at Georgia Tech does a lot of interesting work with larger-scale R/C vehicles. That being said, by adding more degrees of freedom you're making the configuration space of the robot higher. In many cases, higher dimension configuration spaces are more difficult for traditional sampling-based approaches. Learning approaches may be able to bias their solutions to avoid large sections of this configuration space.
The servos are certainly where you can spend most money when building a robot for research. I tried building small research robots with small servos ($25 each), see results at the end of this video: https://m.youtube.com/watch?v=q8jgu-EtCFc
The total cost for one robot was well below 1k. The servos are $25 each (Turnigy 306G if remember correctly), all connecting parts are 3d printed, the electronics and batteries are about $50, and you could add a raspberry pi for $50. So in total <$500, depending on number of DOFs.
Of course, these servos don't have the same torque as Dynamixles, for example. But they are lighter!