Questions about data collection details for reproducing OpenVLA-SFT results

[HUAFOR]

Hi, I'm trying to reproduce the OpenVLA-SFT results on Simpler-Env subject generalization tasks. I have some questions about the data collection process:

1. For in-domain tasks like PutCarrotOnPlate, how many episodes did you collect per task? The paper mentions 100 trajectories total but doesn't specify the per-task breakdown.

2. For SFT training, how many successful episodes did you use per task? Since OpenVLA-SFT uses only successful trajectories, it would be helpful to know the target number.

3. For challenging tasks where episodes often fail (e.g., PutSpoonOnTableClothInScene where all 20 episodes failed in my attempts), what was your approach? Did you:

   • Extend the episode range until getting enough successes?
   • Make environmental adjustments?
   • Have a minimum success rate requirement?

This information would help ensure proper replication of the training setup. Thanks!

[zzj1111]

Hi HUAFOR!

1. The 100 trajectories consists of 25 trajectories per-task for the 4 tasks in Simpler-Env(25*4)
2. For Simpler-Env, we use 100 trajectories(the whole dataset) one time to train the model, and then we eval the model across 4 tasks.
3. Can you tell me more detail about your experiments? It is abnormal for OpenVLA-SFT to fail for all the 20 episodes in "PutSpoonOnTableClothInScene".

[Akila-Ayanthi]

Hi @HUAFOR I have the same question. Were you able to generate enough trajectories?