Question about build-in IL and RL algorithms

[lingxiao-guo]

Hi, impressive work！ Your paper evaluated the tasks with ACT and diffusion policy and some rl methods，but I haven't found those algorithms in this codebase. Do you plan to open this part of code？Since it would be more convenient to adjust or compare these algorithms on bigym. Thanks！

[chernyadev]

Hi @lingxiao-guo!

BiGym is now part of our robot learning framework, robobase. Check it out along with the baseline algorithms: https://github.com/robobase-org/robobase.

[lingxiao-guo]

Thanks a lot！would love to give it a try

[AnnyOrange]

Hi @chernyadev,

I’m really impressed by the quality of the dataset you’ve shared – it’s fantastic! I’ve been running the BiGym tasks using the baseline algorithms you provided, specifically with the diffusion method. The command I used was:

python train.py method=diffusion env=bigym/dishwasher_close env.episode_length=1000 replay.nstep=1

However, after running the training, the results do not seem to complete the specified tasks as expected. I’m not sure if the issue is due to the episode_length being too small or if there are other factors at play. Could you kindly share the command you used when running these tasks? Any insights would be greatly appreciated!

Thank you so much for your help!

[chernyadev]

Hi @AnnyOrange, the problem is related to the fact that demonstrations were collected using the previous version of BiGym. I recommend switching to this branch: https://github.com/chernyadev/bigym/tree/fix_original_demos.

This doesn’t resemble the original functionality completely, but you should end up with higher numbers of successful demonstrations.

[AnnyOrange]

Thanks! This is helpful.

[AnnyOrange]

@chernyadev,Hi!

I am currently encountering an issue while running the act method in robobase (bigym). The error traceback I am getting is shown below:

Upon investigating the code in robobase/robobase/method/act.py, I found that the error occurs during the initialization of output_shape in the build_actor function.

I printed out the value of self.encoder, and it turns out to be None. As a result, the output_shape function is not being called, leading to the error. In the ActBCAgent(BC) class, self.encoder is None by default.

Could you please advise on how to resolve this issue? Should I manually set self.encoder = ImageEncoderACT and call the output_shape function, or is there another underlying issue that I might have overlooked?

[younggyoseo]

Hi, you might want to check PR to Robobase (swirl-uk/robobase#5) that makes ACT work on BiGym (pixels) and CQN-AS's codebase (https://github.com/younggyoseo/CQN-AS) that supports training demo-driven RL agents on BiGym.

[balabala11111]

Hi, @AnnyOrange ! I'm facing the same "Error executing job with overrides" issue as you. Have you resolved it yet? If you have, could you please share how did you fix the issue? Thank you so much!

[AnnyOrange]

Hi @balabala11111! It's been a while since I last ran this code, so I can't recall the exact details of how the issues were resolved. However, I remember that after @younggyoseo updated the robobase (https://github.com/[robobase-org/robobase]) repository, I didn't encounter any further problems with running it. If I recall correctly, to run bigym, you just need to use the command python3 train.py launch=act_pixel_bigym env=bigym/xxx.

[balabala11111]

@AnnyOrange I have solved the problem under the guidance.Thank you soooooo much for your kind reply!