starvla_qwenpiv3_la_pretrain

StarVLA DROID LA QwenPI-v3 Pretrain

This repository contains the StarVLA/QwenPI-v3 Language-Action pretraining checkpoint on DROID LA data.

Checkpoint

• Model: starvla_droid_la_qwenpiv3_pretrain
• Step: 10000
• Training data: DROID LA segments (droid_seg_full_raw, 33,116 episodes)
• Training mode: LA pretraining with vision_masked: true
• Action/state space: DROID 7-DoF EEF (action_dim: 7, state_dim: 7)
• Base VLM: Qwen/Qwen2.5-VL-3B-Instruct

This is a pretraining checkpoint, not a directly evaluated downstream policy. For downstream finetuning on a different action space, reload the compatible StarVLA/QwenPI-v3 modules and reinitialize incompatible action heads as required by your codebase.

Files

checkpoints/steps_10000_pytorch_model.pt
config.yaml
config.full.yaml
dataset_statistics.json

dataset_statistics.json is required for DROID action denormalization. Keep the directory structure unchanged if your training or inference server expects the checkpoint directory layout used by the released config.

Usage Notes

This checkpoint is intended to be loaded with the StarVLA/QwenPI-v3 codebase. The public configs use Qwen/Qwen2.5-VL-3B-Instruct as the base VLM path.

For continued training or evaluation on your own machine, update:

datasets:
  vla_data:
    data_root_dir: ./LA-33K

to your local DROID LA data path.

Citation

If you use this checkpoint, please cite the corresponding StarVLA / LA4VLA work and this model repository.