Bad Performance for doorman on 4090 GPU

[Sivan666666]

Hi, thank you very much for open-sourcing this great project.

I am trying to reproduce the Doorman task. I trained the teacher-stage policy on a single RTX 4090 GPU. After around 40k training iterations, the robot is able to approach the door, turn the handle, and open the door. However, it usually stops in front of the doorway and does not successfully walk through the door.

I did not modify any reward functions or task settings.

Has anyone encountered a similar issue? Are there any recommended hyperparameters, training tricks, or debugging suggestions for improving the final “passing through the doorway” behavior?

Any advice would be greatly appreciated. Thank you!

[ziyu-heha]

Hello, I encountered the same situation while trying to reproduce the results. May I ask if you have made any new progress?

[gunnxx]

Hello, I am wondering how/where did you obtain the HOMIE locomotion models?

[Exentr0]

Hey, I was struggling with the same issue. I ran on a laptop 5090 with 3072 parallel envs, also trained for around 40k iters, and mine actually got to completing the task rarely — so the learning does work, it's just really slow because of the learning rate. The KL-adaptive lr controller pins the lr at its 1e-5 floor from the very start (the desired_kl: 0.005 in the exp config is too tight, and the logged lr metric doesn't show the real lr — you can check checkpoint["args"].learning_rate in any of your checkpoints, mine was stuck at 1e-05).

Running with ++algo.config.desired_kl=0.05 fixed it — the actual lr jumped from 1e-5 to ~2.5e-4 (about 25x larger), so training is no longer crawling at the floor and actually learns properly. That gave roughly a 25x speedup and now the training matches the results in the paper with the experiment config.

[gunnxx]

Hi @Exentr0 , I am sorry for such silly question, I am wondering where did you obtain the HOMIE policy? I could not find it here nor the homie codebase.

[Exentr0]

Hi @gunnxx Not a silly question — the README is confusing here. The HOMIE models are already in this repo via Git LFS (models/model_walk.pt and models/model_stand.pt). If those files are ~130-byte text files after cloning, you're missing the LFS content — run:

git lfs install
git lfs pull

[gunnxx]

@Exentr0 Ohhh I didnt notice that. Thank you so much!