openvla-oft_skill

Overview

This skill fine-tunes and evaluates OpenVLA-OFT and OpenVLA-OFT+ policies for robot action generation with continuous action heads, LoRA adaptation, and FiLM conditioning. It packages workflows for LIBERO simulation and ALOHA real-world setups, including training, evaluation, LoRA merge, and server-client deployment. Use it to reproduce paper results, train custom continuous action heads (L1 or diffusion), and debug cross-GPU or normalization problems.

How this skill works

The skill wraps official OpenVLA-OFT scripts to freeze the VLA backbone and adapt with LoRA adapters while replacing tokenized actions with continuous action heads (L1 regression or diffusion). It supports OFT (2-camera, no FiLM) for LIBERO and OFT+ (3-camera, FiLM) for ALOHA, plus dataset conversion, training launch commands, evaluation runners, and a VLA inference server. It includes utilities for invariant checks (config parity), LoRA merging, and common failure diagnostics (unnormalization keys, GPU type issues).

When to use it

• Reproducing OpenVLA-OFT paper experiments on LIBERO or ALOHA
• Training custom continuous action heads (L1 regression or diffusion) with LoRA adapters
• Deploying a VLA inference server for ALOHA robot evaluation (server-client)
• Debugging normalization, LoRA merge, or cross-GPU checkpoint issues
• Evaluating pretrained OpenVLA-OFT checkpoints on LIBERO or ALOHA

Best practices

• Keep training and inference flags identical (use_film, num_images_in_input, use_proprio, lora_rank, center_crop)
• Prefer LoRA adaptation to full fine-tuning; merge adapters for deployment on different GPU types
• Validate dataset normalization keys and set --unnorm_key to match checkpoint stats
• Use recommended software stack (Python 3.10.14, PyTorch 2.2.0) and route caches to scratch on clusters
• Run intermediate evals (50k, 100k, 150k) and choose checkpoints by task success, not loss

Example use cases

• Run LIBERO evaluation on a pretrained checkpoint to reproduce reported success rates
• Fine-tune OpenVLA-OFT with RLDS LIBERO datasets using torchrun across 8 A100 GPUs and LoRA rank 32
• Train OFT+ on ALOHA with FiLM and 3 image streams, start the VLA server, and run client-side robot rollouts
• Merge LoRA adapters into a deployable checkpoint when moving between GPU types
• Parse LIBERO logs to extract per-task success rates and validate reproduction