search

chernyadev / bigym

Demo-Driven Mobile Bi-Manual Manipulation Benchmark.
https://chernyadev.github.io/bigym/
Apache License 2.0
113 stars 15 forks source link

Question about build-in IL and RL algorithms #30

Open lingxiao-guo opened 1 week ago

lingxiao-guo commented 1 week ago

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 commented 1 week ago

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 commented 1 week ago

Thanks a lot！would love to give it a try

AnnyOrange commented 1 week ago

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 commented 1 week ago

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 commented 1 week ago

Thanks! This is helpful.

AnnyOrange commented 1 week ago

@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:

Error Screenshot

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.

Code Screenshot

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?