The official gitter channel for the participants of the NIPS Learning to Run Challenge
Anyone else from the AI4Prosthetics challenge at AMLD this year ?
======
Announcing the Learning2Move challenge for NeurIPS 2019 : https://www.aicrowd.com/challenges/neurips-2019-learn-to-move-walk-around
Announcing the Learning2Move challenge for NeurIPS 2019 : https://www.aicrowd.com/challenges/neurips-2019-learn-to-move-walk-around
I've been trying to understand what the target velocity map means in (http://osim-rl.stanford.edu/docs/nips2019/environment/), but I'm having a difficult time. I'm not able to understand the terse description given in the linked webpage. Can anyone point me to more information about the velocity map? It looks like it indicates the target velocity, but I don't understand why the target velocity has to be a map (instead of a single vector value).
can we assume that the distribution of the velocity vectors in the global velocity map is constant? I.e. it looks like the velocity fields are a 2d gaussian with constant variance but varying mean (final location). So can we assume that the velocity field will always have the constant "shape" of the gaussian? If that's the case, then I think the velocity map might not be necessary.
Hi, @brokenBrain . A new simulation creates a new velocity map where the final location is generated by a 2d gaussian, but the velocities toward that location varies independently to the final location. In addition, reaching the final location may not be the end of the task ;)
I noticed that many submissions were failing because of the lack of
msgpack installed. I have created an issue (stanfordnmbl/osim-rl#191) and will include it as a dependency in the next release. In the meantime please do add msgpack and msgpack_numpy to your conda environment (pip install msgpack msgpack-numpy) before exporting the environment.yml
Another note is about the
If you add
debug mode.If you add
”debug”:true to your aicrowd.json, then the evaluator will give you access to your evaluation logs if your evaluation fails. This can make debugging cycles much easier.
Hi, I have an issue on setting the project=True and obs_as_dict = True.
When I try locally project=True and obs_as_dict = True, after flattening I get an array of shape (339,)
Sending this for evaluation, I surprisingly get an array of shape (710). Is there an explanation of this or a workaround if possible???
thank you in advance!
When I try locally project=True and obs_as_dict = True, after flattening I get an array of shape (339,)
Sending this for evaluation, I surprisingly get an array of shape (710). Is there an explanation of this or a workaround if possible???
thank you in advance!
Hi, is the simulator on crowdai for evaluation set project=True and obs_as_dict = True as you mentioned (https://www.aicrowd.com/challenges/neurips-2019-learn-to-move-walk-around) in clarification?
Because I get the inverse output during evaluation, like it is set as project=False and obs_as_dict = False
@spMohanty @SeungmoonS_twitter
Because I get the inverse output during evaluation, like it is set as project=False and obs_as_dict = False
@spMohanty @SeungmoonS_twitter
Just submitted the random test controller. Didn't recieve any notification on succesfull submission. Is it normal?
I decided to try out the new submission method, as I have some difficulties with nvidia docker, but I get this error from it: "ServerErrorPrint Wrong client version. Please update to the new version. Read more on https://github.com/stanfordnmbl/osim-rl/docs 400
Wrong client version. Please update to the new version. Read more on https://github.com/stanfordnmbl/osim-rl/docs". I find it unclear on where to go in the docs, and it was ran from a fresh clone from github, so I'm not sure why it would be out of date.
Wrong client version. Please update to the new version. Read more on https://github.com/stanfordnmbl/osim-rl/docs". I find it unclear on where to go in the docs, and it was ran from a fresh clone from github, so I'm not sure why it would be out of date.
@brokenBrain Yes, it is [1.4, 0] for all episodes: https://github.com/stanfordnmbl/osim-rl/blob/55db058307cb60ffdc9852350f912e4bb317c411/envs/target/v_tgt_field.py#L23
Hi,
What?
I just want to share with you my starter kit for this year competition - distributed DDPG agent based on Catalyst.RL framework.
https://github.com/Scitator/learning-to-move-starter-kit
Why I am doing this?
- I am participating in this competition for last 2 year. And I love it, I think, that the competition mission is really important and valuable.
- I already have 2 publicly available solutions (both are 3rd place winners):
2017 - https://github.com/Scitator/Run-Skeleton-Run
2018 - https://github.com/Scitator/neurips-18-prosthetics-challenge - last year I also open-sourced Catalyst - DL&RL framework for fast and reproducible experiments. The RL part was developed during 2018 year competition.
Based on all this I want to boost community activity in this competition and RL overall.
For more RL news and this competition insights - you can find me by twitter.
Thanks!
@unghee_lee_twitter You can check out our last year's osim-rl repository (https://github.com/stanfordnmbl/osim-rl/tree/ver2.1). It is a human model with a passive below-knee prosthesis.
I have an intuition, that full raw observation could be better for agent training.
@Scitator Thanks for sharing your thoughts. You are probably right that it is better for training. However, we designed the current observation dict as it seems closer to biological sensory data used in humans and previous studies (e.g. https://youtu.be/ZkOrRcc4dWg) show that it is possible to control locomotion with those data. We will most likely keep the current setting unless there is some biology-based concern or fundamental limitation in training.