I'm building a bipedal robot inspired by Agility Robotics' Cassie. It uses 10 OpenTorque actuators. 

These actuators are backdrivable and able to absorb kinetic energy during the walk cycle. As a result of this, there's no need for the complex series-elastic linkages present in Cassie. Instead, each leg consists of a simple parallelogram linkage with an extra joint at the ankle. 

Carbon fiber tubes are used for the linkages in order to keep inertia as low as possible. The total weight is in the ballpark of 30 pounds -- substantially lower than the 60-70 pounds that Cassie weighs. This is helped by the use of 3d-printed plastic rather than CNC aluminum parts. (All the structural parts will be printed out of NylonX.)

To control this robot, I'm going to use reinforcement learning. I'll create an OpenAI training environment with a simulation of the robot, then let the controller learn a stable walking gait on its own. This is a lot easier than programming a walking gait by hand, and it's already been done successfully on the Minitaur robot.