Go1's Getup task is not working; when it turns over, it remains prone

[lmtss]

Neither using brax nor rsl-rl allows the Go1 to stand up after turning over; it remains prone.

[Rian-Jo]

for me, i trained several times with checkpoints for Getup task. the robot gets done after few trainings.

[lmtss]

for me, i trained several times with checkpoints for Getup task. the robot gets done after few trainings.

Wow, that's awesome! Approximately how many timesteps or iterations in total?

[lmtss]

for me, i trained several times with checkpoints for Getup task. the robot gets done after few trainings.

My goodness, I did it! I tried training for 750 million timesteps (the default configuration is 50 million), and it worked. Thank you for sharing; this is very helpful.

[kevinzakka]

Hi, getup can be trained extremely efficiently. We only do extra passes with a power curriculum to deploy safer policies on hardware. Can you try using this notebook and tell me if you can reproduce the result?

[Rian-Jo]

@kevinzakka

i have tested fall recovery task.

test.mp4

i just added installation parts from the notebook you shared and ran training on colab. is there something wrong?

[lmtss]

Hi, getup can be trained extremely efficiently. We only do extra passes with a power curriculum to deploy safer policies on hardware. Can you try using this notebook and tell me if you can reproduce the result?

I trained getup directly by running train_jax_ppo.py instead of using the notebook. Could this be the source of the problem? I can try running the notebook to see if it reproduces the result.

[xander-2077]

Hello! Can you show me the reward curve and the video of the training effect? @lmtss ​​I am not sure about my training results. In addition, the effect of this video looks quite different from that in https://playground.mujoco.org/, which seems unreasonable. Have you tried to transfer it to a real robot?