VLA-Adapter-LIBERO-Spatial-15000
VLA-Adapter policy: a Qwen2.5-0.5B prism backbone (fused DINOv2 + SigLIP vision) with a layer-wise Bridge-Attention action head, fine-tuned on LIBERO.
| Backbone | qwen25-0_5b-extra |
| LLM | 896 hidden x 24 layers |
| Vision | vit_large_patch14_reg4_dinov2.lvd142m + vit_so400m_patch14_siglip_224 |
| Image size | 224x224 |
| Action head | MLP-ResNet, Bridge Attention (Pro) |
| Objective | L1 regression over the action chunk |
| 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--15000_chkpt |
| unnorm_key | libero_spatial_no_noops |
| Evaluated on | libero_spatial |
| Success rate | 98.6% |
Component files were renamed to the convention the evaluation loader expects:
action_head--15000_checkpoint.pt->action_head--checkpoint.ptproprio_projector--15000_checkpoint.pt->proprio_projector--checkpoint.pt
Evaluating
The evaluation code loads the action head and proprio projector from a local directory, and only accepts a Hub repo id if it is in its hardcoded allowlist, so download the repo first:
hf download Zenma/VLA-Adapter-LIBERO-Spatial-15000 --local-dir ckpts/VLA-Adapter-LIBERO-Spatial-15000
python experiments/robot/libero/run_libero_eval.py \
--pretrained_checkpoint ckpts/VLA-Adapter-LIBERO-Spatial-15000 \
--task_suite_name libero_spatial \
--num_images_in_input 2 \
--use_proprio True \
--use_l1_regression True \
--num_open_loop_steps 8
use_pro_version is inferred from whether the checkpoint path contains "Pro",
so keep that substring in the directory name -- this checkpoint is a Pro head.
Loading the VLM alone
from transformers import AutoModelForVision2Seq, AutoProcessor
processor = AutoProcessor.from_pretrained("Zenma/VLA-Adapter-LIBERO-Spatial-15000", trust_remote_code=True)
vla = AutoModelForVision2Seq.from_pretrained("Zenma/VLA-Adapter-LIBERO-Spatial-15000", trust_remote_code=True)
This gives the backbone only. Action prediction additionally needs
action_head--checkpoint.pt and proprio_projector--checkpoint.pt from this
repo, instantiated as in experiments/robot/openvla_utils.py.