Zenma/VLA-Adapter-LIBERO-Spatial-5000

1---2library_name: transformers3pipeline_tag: robotics4license: mit5tags:6- robotics7- vision-language-action8- vla9- libero10- openvla11- vla-adapter12---13 14# VLA-Adapter-LIBERO-Spatial-500015 16VLA-Adapter policy: a Qwen2.5-0.5B prism backbone (fused DINOv2 + SigLIP vision)17with a layer-wise Bridge-Attention action head, fine-tuned on LIBERO.18 19| | |20|---|---|21| Backbone | qwen25-0_5b-extra |22| LLM | 896 hidden x 24 layers |23| Vision | vit_large_patch14_reg4_dinov2.lvd142m + vit_so400m_patch14_siglip_224 |24| Image size | 224x224 |25| Action head | MLP-ResNet, Bridge Attention (Pro) |26| Objective | L1 regression over the action chunk |27| Checkpoint | configs+libero_spatial_no_noops+b16+lr-0.0001+lora-r64+dropout-0.0--image_aug--VLA-Adapter--libero_spatial_no_noops--20260727_221630--5000_chkpt |28| unnorm_key | `libero_spatial_no_noops` |29| Evaluated on | libero_spatial |30 31Component files were renamed to the convention the evaluation loader expects:32 33- `action_head--5000_checkpoint.pt` -> `action_head--checkpoint.pt`34- `proprio_projector--5000_checkpoint.pt` -> `proprio_projector--checkpoint.pt`35 36## Evaluating37 38The evaluation code loads the action head and proprio projector from a local39directory, and only accepts a Hub repo id if it is in its hardcoded allowlist,40so download the repo first:41 42```bash43hf download Zenma/VLA-Adapter-LIBERO-Spatial-5000 --local-dir ckpts/VLA-Adapter-LIBERO-Spatial-500044 45python experiments/robot/libero/run_libero_eval.py \46    --pretrained_checkpoint ckpts/VLA-Adapter-LIBERO-Spatial-5000 \47    --task_suite_name libero_spatial \48    --num_images_in_input 2 \49    --use_proprio True \50    --use_l1_regression True \51    --num_open_loop_steps 852```53 54`use_pro_version` is inferred from whether the checkpoint path contains "Pro",55so keep that substring in the directory name -- this checkpoint is a Pro head.56 57## Loading the VLM alone58 59```python60from transformers import AutoModelForVision2Seq, AutoProcessor61 62processor = AutoProcessor.from_pretrained("Zenma/VLA-Adapter-LIBERO-Spatial-5000", trust_remote_code=True)63vla = AutoModelForVision2Seq.from_pretrained("Zenma/VLA-Adapter-LIBERO-Spatial-5000", trust_remote_code=True)64```65 66This gives the backbone only. Action prediction additionally needs67`action_head--checkpoint.pt` and `proprio_projector--checkpoint.pt` from this68repo, instantiated as in `experiments/robot/openvla_utils.py`.69