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

crowdAI is shutting down - please read our blog post for more information

NIPS 2017: Learning to Run

Reinforcement learning environments with musculoskeletal models


Pelvis trial threshold

Posted by Anton_Pechenko over 4 years ago

Currently we have trials that ends if the pelvis of the model goes below 0.65 meters. Maybe it is reasonable to lower that limit to 0.5 or 0.4 meters? Is 0.65 is too high? Sometimes I get sudden trial end while my agent is look like moving correctly.

Also do you show the shortest try from submission? Why not to show the longest?


Posted by spMohanty  over 4 years ago |  Quote

@Anton Pechenko : I will let @Lukasz answer the first question about the threshold height for the end of a simulation. I personally feel that 0.65 is a reasonable number , but then again it is indeed very subjective.

Regarding the visualisations, indeed, that is a good point.
Currently we show the visualisation of just the first trial. We are working on updating the visualisations to all the three trials. It should be deployed within the week.

Cheers, Mohanty

Posted by Adam_Stelmaszczyk  over 4 years ago |  Quote


If I may suggest something, I think it will be better for everyone to first speed-up the loading time of the leaderboard by displaying GIFs or adding pagination (or “more” button) and then show 3x more visualizations :)


Right now my browser is crashing when displaying the Leaderboard, because of 102 large MP4 files ;_;


Posted by Yongjin  over 4 years ago |  Quote

Regarding the pelvis trial threshold, I agree with Anton.

My agent makes a big jump over an obstacle and it seems that it successfully lands on the ground,

but while it tries to recover its pose, a simulation ends because the pelvis height is below the threshold.

The agent’s landing is not bad considering the size of obstacles and the impact after jumping.

In reality, we also need to lower our knees to damp an impact after a big jump.