logicvla-7b-libero-spatial

OpenVLA-7B Fine-tuned on LIBERO-Spatial

This repository contains a fine-tuned OpenVLA-7B checkpoint for LIBERO-Spatial.

The model uses OpenVLA-7B as the backbone and builds upon the OpenVLA-OFT training framework with additional training-time supervision and optimization components.

Additional methodological details, code, and citation information will be released after the review process.

Quick Start

This example demonstrates how to load the checkpoint for LIBERO-Spatial evaluation.

```python import pickle

from experiments.robot.libero.run_libero_eval import GenerateConfig from experiments.robot.openvla_utils import get_action_head, get_processor, get_proprio_projector, get_vla, get_vla_action, from prismatic.vla.constants import NUM_ACTIONS_CHUNK, PROPRIO_DIM

cfg = GenerateConfig( pretrained_checkpoint="zijie-ai/logicvla-7b-libero-spatial", use_l1_regression=True, use_diffusion=False, use_film=False, num_images_in_input=2, use_proprio=True, load_in_8bit=False, load_in_4bit=False, center_crop=True, num_open_loop_steps=NUM_ACTIONS_CHUNK, unnorm_key="libero_spatial_no_noops", )

Load VLA policy and processor

vla = get_vla(cfg) processor = get_processor(cfg)

Load continuous-action prediction head

action_head = get_action_head(cfg, llm_dim=vla.llm_dim)

Load proprioceptive-state projector

proprio_projector = get_proprio_projector(cfg, llm_dim=vla.llm_dim, proprio_dim=PROPRIO_DIM)

Load a LIBERO observation

with open("experiments/robot/libero/sample_libero_spatial_observation.pkl", "rb") as file: observation = pickle.load(file)

Generate an action chunk

actions = get_vla_action(cfg, vla, processor, observation, observation["task_description"], action_head, proprio_projector)

print("Generated action chunk:") for act in actions: print(act) ```

Related Checkpoints

An intermediate 15K-step checkpoint is also available:

• 15K checkpoint