OpenVLA

OpenVLA supplies vision-language-action model code and checkpoints for image/instruction-conditioned manipulation. Its interface predicts actions, while a separately implemented robot controller must interpret and execute them. The repository also links later fine-tuning/tokenization projects; their performance claims are not automatically features or measurements of this selected code/checkpoint. [1] [8]

The reference environment uses Python 3.10, with package metadata allowing 3.8 or later and pinning important dependencies including torch 2.2.0 and transformers 4.40.1. Minimal inference and full training installations differ. The README example moves inputs/model to CUDA and uses bfloat16, with FlashAttention optional. A CPU selection branch in a deployment script would not prove a compatible CPU inference path. No environment or model was loaded here. [1] [5] [6] [8]

A concrete offline inspection route is the openvla-7b processor/model interface and vla-scripts/deploy.py. The README's Bridge example uses unnorm_key="bridge_orig" to restore the intended action scale. Its camera and robot.act calls are placeholders for real interfaces, not a complete device program. The model card describes seven action components: end-effector deltas (x, y, z, roll, pitch, yaw) and gripper control. Per-robot, per-dataset unnormalization statistics are required; the inspected evidence does not establish universal action units. Match observation framing and the actual robot domain before interpreting predictions. [1] [7] [8]

The README says new domains generally need fine-tuning with target-robot demonstrations; out-of-the-box behavior should not be assumed for a different embodiment. The LoRA route uses vla-scripts/finetune.py, accepting RLDS data registered with the dataset configuration and transform or a custom PyTorch Dataset wrapper. Full fine-tuning instead uses the Prismatic-compatible openvla-7b-prismatic checkpoint with vla-scripts/train.py; its output must be converted for Hugging Face inference using vla-scripts/extern/convert_openvla_weights_to_hf.py. The LIBERO evaluation examples require their separately matched environment and checkpoints. The upstream LoRA example uses an A100 80 GB and reports approximately 72 GB at batch size 16 with gradient accumulation 1, or approximately 27 GB with smaller batches. Full fine-tuning requires substantially more compute. These are upstream training figures, not measured inference minima or current GPU purchase recommendations. [1]

The code is MIT. The official pinned model card also declares MIT for checkpoints and identifies a Llama-2 backbone, while the code README separately warns that pretrained models inherit underlying Llama Community License terms. Preserve those layered declarations rather than offering blanket weight clearance or assuming the notices are necessarily irreconcilable. Dataset contributions and external backbones retain their own terms. [1] [2] [7]

Only model-card/metadata and selected implementation/setup text were inspected. No weight download/deserialization, prediction, HTTP deployment, fine-tuning, benchmark, camera connection or robot actuation was performed. Begin with offline input/action interpretation and a selected compatible checkpoint before any hardware integration. [7] [8]