Open student-petercai opened 1 year ago
Hi @student-petercai ,
The policy only goes forward, but it should be able to turn both left and right. If it does not turn right, you might have a problem in the joystick.
If you want to specialize to the Go1, re-train a policy using this code: https://github.com/antonilo/rl_locomotion/tree/master.
Specifically, you need to add the urdf of the GO1 in here: https://github.com/raisimTech/raisimLib/tree/master/rsc/a1.
If you want it to go backwards, change the randomization such that max_speed is also negative. Rember also to add a special case for the reward when velocity is negative here. Finally, if max_speed<0, set useRef=False (so it will not use reference motion from the base policy).
I hope this helps! If you have a working policy, feel free to share it!
Hi @antonilo,
Thank you for the informative reply! We will let you know when we have a good policy for Go1.
Currently, we are also training the vision predictor and have noticed that the images captured from our D435i are quite blurry, especially when climbing up stairs due to shaky movements. I have attached two sample blurry images below. I am wondering if you have experienced a similar issue and if it is okay to train with such images. Additionally, you mentioned that you uploaded some dummy datasets in the repository, but we were unable to locate them. Did you forget to upload them?
Also, would you advise on how many images are required to train a good vision policy?"
Thank you!
Hi @student-petercai,
It is totally normal that images are blurry. But this is not a problem since the network is trained on this data. This is the beauty of continual real-world learning! If you want to decrease the blur, consider increasing the strength of the camera mount (if you're using the one I shared). I have another version that is a bit stronger that I am happy to share if you need it.
I am still figuring out some details for sharing data, will do that asap.
Regarding data collection, you should already see quite good performance in around 30min/1hr of walking data. If you're also walking backward, estimating with a forward camera might be challenging. Please feel free to share the code for the go1 that you're using if you'd like. We can make a separate branch and release.
Hi @student-petercai, Any progress on the Go1 ? I'd love to look at the modifications you've made ! Thank you.
Hi,
Thank you for sharing the code. I'm currently testing the blind policy on my Unitree Go1 robot, and I made a few modifications to get it up and running. However, I have a couple of questions:
1) I noticed that I can only remote control the robot to move forward or turn left, but I cannot move it in reverse. Is this normal, or is it an issue of a difference between A1 and Go1?
2) If I want to retrain the policy on the Go1, could you advise me on the major adaptations I would need to make?
Thank you!