In the curriculum 2 (forward jumping),how did you push the robot torward the goal?

[jassonye]

i am quite confused about the setting of the command velocity depanding on the relation between joint friction and flight time. and i am also curiousity about the push velocity is only on X axis and Y axisA or also on Z axis. Really thanks if you could help me figure out, i try so many times but always can not get a good result in trainning of forward jumping.

[vassil-atn]

Hi, the push velocity depends on the target distance divided by the flight time. We collected a simple dataset of flight time vs joint friction from the upwards jump stage and fit a line through it (that's the a and b on lines 1133-1134 in legged_robot.py). It's not necessary to do that - you can also choose a constant flight time, but that will tend to overshoot for small friction and undershoot for large friction. The push velocity is only on the planar motion (XY), not on the Z (line 1696).

I recommend disabling or lowering the joint friction/damping initially to test if your policy can learn the behaviour well.

[jassonye]

hi!Thank you for your respond. I will try it.but i also have some question, cause i changed some env argument, and i found the robot will not jump forward, although the figure of means rewards tells good,thats because the most of curriculum level of command stay at 0,and the mean curriculum is about 0.01 or so,whcih means the jump distance is really short, and i also use the original argument (download from the github) to train the policy and i find the mean curriculum is also near to zero, could you please help me about it? Besides, i think reset_landing_error = 0.2 # [in %/100] (line 96 in go1_cfg.py) means the distance between the real landing point and the desire landing point instead of percentage error. and one more thing btw,is it reasonable to set the success rate = 0 as the standard of level move down/up. Really appreciate!