[pbmonster]

The jumps are pretty impressive, this thing has some power. I'd be very curious how fast you could get this dog with some reinforcement learning for a proper transverse gallop gait - and if it converges towards a gallop naturally, or if it discovers some other fast gait patterns during learning.

Depending on the max speed of the motors/legs, giving it longer foot pads might be necessary for a good gallop. Intuitively, it looks a bit... "low gear" in the videos.

[SamBam]

I wonder how accurate a virtual model could be made of this, which could be iterated on millions of times faster.

My first project as a research assistant in AI was doing evolutionary algorithms on Khepera robots, which had a virtual Java implementation. We were able to evolve some pretty cool behaviors, although I don't know what would have happened if we had uploaded them into a physical Khepera robot.

[pbmonster]

I'd expect pretty good results from simulation until you start bending/deforming the limbs or until the foot pads start slipping.

I also wonder if you can read the impact force from the legs with the ground from the motor controllers (should be able to infer that from motor current + leg displacement). Learning to gallop without "feeling" the legs impacting sounds extremely difficult.