fine-tuning-openvla-oft - Fine-tune and evaluate OpenVLA-OFT robot policies
Fine-tunes and evaluates OpenVLA-OFT policies with continuous action heads, LoRA adaptation, and FiLM conditioning for LIBERO and ALOHA.
Tags
Updated: 2026-09-29OpenVLAOpenVLA-OFTVLARoboticsFine-TuningLIBEROALOHALoRAFiLMAction ChunkingDeploymentContinuous Actions
Capabilities
What this skill does
- Fine-tune OpenVLA-OFT policies
- Evaluate LIBERO checkpoints
- Train continuous action heads
- Adapt models with LoRA
- Apply FiLM conditioning
- Merge LoRA weights
- Parse evaluation logs
- Deploy server-client inference
Inputs
- OpenVLA-OFT repository
- Pretrained model checkpoints
- RLDS training datasets
- ALOHA demonstrations
- Task suite configuration
- W&B project credentials
Outputs
- Fine-tuned model checkpoints
- LIBERO success rates
- ALOHA inference results
- Merged LoRA model files
- Evaluation log summaries
- Server-client inference state
Requirements
- Python 3.10.14 environment
- NVIDIA GPU resources
- PyTorch 2.2.0
- PEFT 0.11.1
- Transformers 4.40 or newer
- FlashAttention2 installation
- Eight GPUs for paper-scale training
- Custom Transformers fork for reproduction