VQVLA quantized LIBERO weights

GPTVQ archives used by the mixed execution-state 4bit / transition-state 3bit inference path in ACA-Lab/VQVLA. These are inference-only quantization archives, not NVIDIA's original LIBERO checkpoints. Obtain the corresponding LIBERO checkpoint separately.

The Isaac-GR00T archives are stored below the Isaac-GR00T/ subdirectory so this repository can also host weights for other VLA models. From the Isaac-GR00T project root, download that subdirectory and expose it as quant_weight/:

git lfs install
GIT_LFS_SKIP_SMUDGE=1 git clone --depth 1 --filter=blob:none --sparse \
  https://huggingface.co/LeoJiang123/VQVLA .vqvla-weights
git -C .vqvla-weights sparse-checkout set Isaac-GR00T
git -C .vqvla-weights lfs pull --include='Isaac-GR00T/*'
ln -s .vqvla-weights/Isaac-GR00T quant_weight

For the recommended mixed configuration, model1 uses the execution-state 4bit archive and model2 uses the transition-state 3bit archive; pass --routing-threshold 0.55 to the server. The archive selected depends on the LIBERO checkpoint directory:

Checkpoint	execution-state 4bit model1	transition-state 3bit model2
libero_spatial	gr00t_spatial_gptvq_4bit_b256_k256.pt	gr00t_spatial_gptvq_hdiag_3bit_b128x128_cb64_v2.pt
libero_object	gr00t_object_gptvq_hdiag_4bit_b256_k256.pt	gr00t_object_gptvq_hdiag_3bit_b128x128_cb64_v2.pt
libero_goal	gr00t_goal_gptvq_hdiag_4bit_b256_k256.pt	gr00t_goal_gptvq_hdiag_3bit_b128x128_cb64_v2.pt
libero_10	gr00t_10_gptvq_hdiag_4bit_b256_k256.pt	gr00t_10_gptvq_hdiag_3bit_b128x128_cb64_v2.pt

OpenVLA-OFT

The OpenVLA-OFT GPTVQ archives for all four LIBERO suites are stored in OpenVLA-OFT/<suite>/ subpaths. They quantize all 1,002 floating-point parameters in the matching fine-tuned model components (VLM, action head, and proprio projector); obtain the original fine-tuned checkpoint separately. Run the following from the OpenVLA-OFT project root to download a suite's pair into .vqvla-weights/:

LIBERO-Spatial

hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-Spatial/openvla_oft_libero_spatial_4bit_b256x256_k256_v3.pt \
  --local-dir .vqvla-weights
hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-Spatial/openvla_oft_libero_spatial_3bit_b128x128_k64_v3.pt \
  --local-dir .vqvla-weights

The tested mixed threshold is 0.79: 98/100 success, with 741/1,369 policy queries (54.13%) routed to transition-state 3bit.

LIBERO-Object

hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-Object/openvla_oft_libero_object_4bit_b256x256_k256_v3.pt \
  --local-dir .vqvla-weights
hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-Object/openvla_oft_libero_object_3bit_b128x128_k64_v3.pt \
  --local-dir .vqvla-weights

The tested mixed threshold is 0.00: 98/100 success, with 1,726/1,826 policy queries (94.52%) routed to transition-state 3bit.

LIBERO-Goal

pip install -U huggingface_hub
hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-Goal/openvla_oft_libero_goal_4bit_b256x256_k256_v3.pt \
  --local-dir .vqvla-weights
hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-Goal/openvla_oft_libero_goal_3bit_b128x128_k64_v3.pt \
  --local-dir .vqvla-weights

Use the downloaded archives at .vqvla-weights/OpenVLA-OFT/LIBERO-Goal/ for the execution-state 4bit and transition-state 3bit model arguments.

The tested LIBERO-Goal mixed configuration uses routing threshold 0.55 and scored 96/100 episodes (96.0%). It routed 956 of 1,581 policy queries to transition-state 3bit (60.47%).

LIBERO-10

The OpenVLA-OFT LIBERO-10 archives are stored under OpenVLA-OFT/LIBERO-10/. They quantize the complete model components (VLM, action head, and proprio projector); obtain the matching fine-tuned checkpoint separately. From the OpenVLA-OFT project root, download both with:

hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-10/openvla_oft_libero_10_4bit_b256x256_k256_v3.pt \
  --local-dir .vqvla-weights
hf download LeoJiang123/VQVLA \
  OpenVLA-OFT/LIBERO-10/openvla_oft_libero_10_3bit_b128x128_k64_v3.pt \
  --local-dir .vqvla-weights

The tested mixed configuration uses threshold 0.00 and scored 92/100 episodes (92.0%), versus 91/100 for the unquantized baseline. It routed 3,617/3,717 policy queries (97.31%) to transition-state 3bit. Threshold 0.00 maximizes transition-state routing for the nonnegative router metric.