Loading
Cookies help us deliver our services. By using our services, you agree to our use of cookies. Learn more

NIPS 2017: Learning to Run

Reinforcement learning environments with musculoskeletal models


Completed
2154
Submissions
623
Participants
85759
Views

Video clips on the leaderboard feels foreign

Posted by QinYongliang over 1 year ago

on latest submission i scored >15 points each run. but the leaderboard video is showing me falling to the ground. shouldn’t it be a bug or something?

2

Posted by spMohanty  over 1 year ago |  Quote

Hi @Qin Yongliang,

The grader runs on difficulty=2 for a total of 3 times with 3 different random seeds. The video on the leaderboard is just from the first of the 3 simulations.

So you can still get an average score of ~15 if your model does poorly on the first run, and well on the second two runs. The best models though would perform similarly across all the three different runs, as in the second round of the challenge, all the top participants will be evaluated on a larger set of runs initialised with different seeds than now.

Regarding the video, we will soon be updating the leaderboard to include simulations of all the three runs, and not just one.

Cheers, Mohanty

Posted by TangDynasty  over 1 year ago |  Quote

Hi @spMohanty

I cannot find any information about second round of the challenge. What’s the schedule and what’s the exact rules for the second round?

Thanks!

Posted by QinYongliang  over 1 year ago |  Quote

@spMohanty I know this part very well. However, the fact is that I accumulated the total reward of all 3 submission on the client side while submitting. they are all greater than 15.

And it’s not just me; everybody on the top of the chart, their submission seemed very wrong. If they scored, as the video suggested, say 5-7 points in the first trial, it would not be possible to average it out and score >15 points in the next two trials.

And from the video you can see that, the specific seeded trial is rather easy (with very small obstacles), and the agents all performed stupidly.

From my 10+ years of experience of programming, this is definitely a bug. The ids of the videos must have been messed up.

Posted by QinYongliang  over 1 year ago |  Quote

after reading the code, I found the problem: https://github.com/kidzik/osim-rl-grader/blob/master/worker_dir/simulate.py#L32-L34

this is unacceptable, since you assume the intergrator of the simulator is PERFECT with UNLIMITED ACCURACY. as an amateur computer scientist, I know this is clearly not true.

I’ll file an issue on the repo.

Posted by spMohanty  over 1 year ago |  Quote

@Qin Yongliang,

++++++++++++++++++++++++

after reading the code, I found the problem: https://github.com/kidzik/osim-rl-grader/blob/master/worker_dir/simulate.py#L32-L34 this is unacceptable, since you assume the intergrator of the simulator is PERFECT with UNLIMITED ACCURACY. as an amateur computer scientist, I know this is clearly not true. ++++++++++++++++++++++++

That is unfortunately by design. A more detailed explanation on this issue can be found on this github issue : https://github.com/stanfordnmbl/osim-rl/issues/52