This video shows a potential solution by locking the center of mass into the area of the foot platform. Without it the robot will fall over most of the time. Not sure how to code it but it definitely looks like it’s in the direction we want to take things. Definitely closer than I was in 2014 when I first got into building my own robots.
Great @new2bots !
@Thot, I’ll try to make a video but I have to go back to RL’s code and the training for several days…
In fact, the robot can fall in many different ways. The principle of Reinfocment Learning is such that at each delta t, it measures the achievement of the goal: having a positive speed in x, with a pelvis height constant while minimizing its energy consumption. For this, it uses a reward function giving a note. We use local rewards (small dt) and larger rewards. Then we apply a new set of actions and redo the loop. So, it all depends on where it is in the temporal evolution, it can fall in front, behind, on the sides… The approach I want to take now is a little different, using a known motion frame and adapting, with that same RL technique with supervised approch, to keep the balance.
I look forward to hearing your progress. My robot building is based upon standing on the backs of you giants who are far more comfortable with coding and testing.
My goal with my poppy chassis is to make a robot with the ability to climb stairs. Mostly I plan to use tank tracks in the feet for quicker horizontal surface navigation. My robot requirements were AI interactions, travel alongside people and ability to climb the stairs. A legged robot made the most sense to me at the time for for stable stair climbing. It’s was my then 6yr old who said “Dad, what if we put treads on it like Wall-E” then SPR-B1TS was shelved 
I wish you the best for the training. Very interested with your results. Poppy has been built at the origin to make this kind of research on “How to learn to walk ?”
By my side, I ask you the type of falling because it is really linked to the physics of the robot (mass distribution, foot contact quality…) and then gives some clues about the reality gap (from simulation to real life)
Here is a prototype I have to study this with two force sensors below the feet and an IMU in each tight (dynamixel motor)
the final goal is the insert this “walking kit” on my Poppy humanoid.