fine-tuning-openvla-oft（经常微调 openvla）

OpenVLA-OFT

Fine-tuning and evaluation workflows for OpenVLA-OFT and OpenVLA-OFT+ from the official openvla-oft codebase. Covers blank-machine setup plus LoRA-based adaptation of OpenVLA for robot action generation with continuous action prediction heads.

Quick start

Clone the public repo, follow the official setup, then evaluate a pretrained LIBERO checkpoint:

git clone https://github.com/moojink/openvla-oft.git
cd openvla-oft
python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint moojink/openvla-7b-oft-finetuned-libero-spatial \
  --task_suite_name libero_spatial \
  --center_crop True \
  --num_trials_per_task 50 \
  --seed 7

Core concepts

What OpenVLA-OFT changes: Standard OpenVLA tokenizes continuous actions into discrete bins, losing precision. OFT replaces this with dedicated continuous action heads (L1 regression or diffusion) while keeping the VLA backbone frozen and adapting via LoRA.

OFT vs OFT+ variants:

Key architecture choices:

• LoRA adaptation: Rank-32 LoRA on VLA backbone (no full fine-tuning needed)
• Continuous actions: L1 regression head (default) or diffusion head
• FiLM conditioning: Feature-wise Linear Modulation for stronger language grounding in OFT+
• Multi-image input: Configurable 2 or 3 camera streams via num_images_in_input

Compute requirements

Expected performance benchmarks

Official results (paper setup, seed=7, 50 trials per task):

Reproduction notes: results are tied to Python 3.10.14, PyTorch 2.2.0, NVIDIA A100, and custom Transformers fork.

When to use vs alternatives

Use OpenVLA-OFT when:

• The target task is robot action generation with visual and language conditioning
• LoRA-based adaptation of openvla/openvla-7b is preferred
• You need official LIBERO or ALOHA workflows from the OpenVLA-OFT paper
• You want continuous action heads (L1 regression or diffusion) instead of tokenized actions

Use alternatives when:

• You need a different VLA architecture (use fine-tuning-serving-openpi for pi0/pi0.5 models)
• You need the NVIDIA Cosmos Policy stack (use evaluating-cosmos-policy)
• You need general LLM fine-tuning without robot action heads

Workflow 1: Set up environment

Copy this checklist and track progress:

Setup Progress:
- [ ] Step 1: Create conda env and install PyTorch
- [ ] Step 2: Install openvla-oft package in editable mode
- [ ] Step 3: Install FlashAttention2
- [ ] Step 4: Verify critical versions

Step 1: Create conda env and clone repo

conda create -n openvla-oft python=3.10 -y
conda activate openvla-oft
git clone https://github.com/moojink/openvla-oft.git
cd openvla-oft
pip3 install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0
pip3 install robosuite==1.4.0

Step 2: Install package

pip install -e .

Step 3: Install FlashAttention2

pip install packaging ninja
pip install "flash-attn==2.5.5" --no-build-isolation

Step 4: Verify versions

import torch, transformers, peft
print(f"PyTorch: {torch.__version__}")         # Expected: 2.2.0
print(f"Transformers: {transformers.__version__}")
print(f"PEFT: {peft.__version__}")             # Expected: 0.11.1

Workflow 2: Evaluate pretrained checkpoints on LIBERO

LIBERO Eval Progress:
- [ ] Step 1: Install LIBERO dependencies
- [ ] Step 2: Choose checkpoint and task suite
- [ ] Step 3: Run evaluation
- [ ] Step 4: Parse and validate results

Step 1: Install LIBERO

git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
pip install -e LIBERO
pip install -r experiments/robot/libero/libero_requirements.txt

Step 2: Choose checkpoint

Step 3: Run evaluation

python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint moojink/openvla-7b-oft-finetuned-libero-spatial \
  --task_suite_name libero_spatial \
  --center_crop True \
  --num_trials_per_task 50 \
  --seed 7

Step 4: Parse results

import re

def parse_libero_log(log_path):
    """Extract per-task success rates from LIBERO eval log."""
    with open(log_path) as f:
        content = f.read()
    matches = re.findall(r"Task (.+?): (\d+)/(\d+) successes", content)
    for task, successes, trials in matches:
        rate = int(successes) / int(trials)
        print(f"  {task}: {rate:.0%} ({successes}/{trials})")

parse_libero_log("experiments/logs/latest.log")

Workflow 3: Fine-tune on LIBERO

Detailed reference: See references/libero-workflow.md for the full LIBERO setup, checkpoint selection strategy, and LoRA merge instructions.

LIBERO Fine-Tune Progress:
- [ ] Step 1: Prepare RLDS dataset
- [ ] Step 2: Launch torchrun with OFT defaults
- [ ] Step 3: Evaluate intermediate and final checkpoints
- [ ] Step 4: Merge LoRA for deployment if needed

Step 1: Dataset

Use RLDS datasets: libero_spatial_no_noops, libero_object_no_noops, libero_goal_no_noops, libero_10_no_noops.

Step 2: Launch training

torchrun --standalone --nnodes 1 --nproc-per-node 8 vla-scripts/finetune.py \
  --vla_path openvla/openvla-7b \
  --data_root_dir /PATH/TO/RLDS/DATASETS/ \
  --dataset_name libero_spatial_no_noops \
  --run_root_dir /YOUR/CHECKPOINTS/ \
  --use_l1_regression True \
  --use_diffusion False \
  --use_film False \
  --num_images_in_input 2 \
  --use_proprio True \
  --batch_size 8 \
  --learning_rate 5e-4 \
  --num_steps_before_decay 100000 \
  --max_steps 150005 \
  --save_freq 10000 \
  --save_latest_checkpoint_only False \
  --image_aug True \
  --lora_rank 32 \
  --wandb_entity YOUR_WANDB_ENTITY \
  --wandb_project YOUR_WANDB_PROJECT

Step 3: Evaluate checkpoints

Evaluate 50k, 100k, and 150k checkpoints — LIBERO-Goal may peak earlier than other suites. Keep best checkpoint per suite by actual task success, not only training loss.

Step 4: Merge LoRA

python vla-scripts/merge_lora_weights_and_save.py \
  --base_checkpoint openvla/openvla-7b \
  --lora_finetuned_checkpoint_dir /PATH/TO/CHECKPOINT_DIR

Workflow 4: Train and evaluate OpenVLA-OFT+ on ALOHA

Detailed reference: See references/aloha-workflow.md for the full ALOHA server-client setup, data preprocessing, dataset registration, and troubleshooting.

ALOHA Progress:
- [ ] Step 1: Preprocess raw ALOHA demonstrations
- [ ] Step 2: Convert to RLDS and register dataset configs
- [ ] Step 3: Fine-tune OFT+ with FiLM and 3 images
- [ ] Step 4: Start VLA server on GPU machine
- [ ] Step 5: Run client-side robot evaluation

Step 1: Preprocess raw data

python experiments/robot/aloha/preprocess_split_aloha_data.py \
  --dataset_path /path/to/aloha_raw/task_name/ \
  --out_base_dir /path/to/aloha_preprocessed/ \
  --percent_val 0.05

Step 2: Register RLDS dataset

Add entries in:

• prismatic/vla/datasets/rlds/oxe/configs.py
• prismatic/vla/datasets/rlds/oxe/transforms.py
• prismatic/vla/datasets/rlds/oxe/mixtures.py

Set ALOHA constants in prismatic/vla/constants.py:

# Expected defaults for ALOHA
NUM_ACTIONS_CHUNK = 25        # Match control frequency (25 Hz)
ACTION_DIM = 14               # 7 joints x 2 arms
PROPRIO_DIM = 14
ACTION_PROPRIO_NORMALIZATION_TYPE = "BOUNDS"  # Absolute joint angles

Step 3: Fine-tune OFT+

torchrun --standalone --nnodes 1 --nproc-per-node 8 vla-scripts/finetune.py \
  --vla_path openvla/openvla-7b \
  --data_root_dir /PATH/TO/RLDS/DATASETS/ \
  --dataset_name aloha_task_name \
  --run_root_dir /YOUR/CHECKPOINTS/ \
  --use_l1_regression True \
  --use_diffusion False \
  --use_film True \
  --num_images_in_input 3 \
  --use_proprio True \
  --batch_size 4 \
  --learning_rate 5e-4 \
  --num_steps_before_decay 50000 \
  --max_steps 100005 \
  --use_val_set True \
  --val_freq 10000 \
  --save_freq 10000 \
  --lora_rank 32

Step 4: Start VLA server (GPU machine)

python vla-scripts/deploy.py \
  --pretrained_checkpoint /PATH/TO/FINETUNED/CHECKPOINT/ \
  --use_l1_regression True \
  --use_film True \
  --num_images_in_input 3 \
  --use_proprio True \
  --center_crop True \
  --unnorm_key aloha_task_name

Server listens on http://<server-ip>:8777/act.

Step 5: Run client evaluation

python experiments/robot/aloha/run_aloha_eval.py \
  --center_crop True \
  --num_open_loop_steps 25 \
  --use_vla_server True \
  --vla_server_url http://<SERVER_IP>:8777 \
  --num_rollouts_planned 50 \
  --max_steps 1500

Critical invariants

These flags must be consistent between training and inference. Mismatches cause silent failures:

Quick validation:

# Verify config parity before long eval runs
train_flags = {"use_film": False, "num_images": 2, "use_proprio": True, "lora_rank": 32}
eval_flags  = {"use_film": False, "num_images": 2, "use_proprio": True, "lora_rank": 32}
for k in train_flags:
    assert train_flags[k] == eval_flags[k], f"Mismatch: {k}: {train_flags[k]} vs {eval_flags[k]}"
print("All flags consistent")

Common issues

Issue: Action quality drops after moving checkpoints across GPU types

Fix: re-merge LoRA adapter on the downstream device:

python vla-scripts/merge_lora_weights_and_save.py \
  --base_checkpoint openvla/openvla-7b \
  --lora_finetuned_checkpoint_dir /PATH/TO/CHECKPOINT_DIR

Issue: Wrong action scale or failed un-normalization

Fix: check --unnorm_key matches dataset statistics in checkpoint:

import torch
ckpt = torch.load("checkpoint/model.pt", map_location="cpu")
print("Available norm keys:", list(ckpt.get("norm_stats", {}).keys()))

Issue: Eval success unexpectedly low

Fix: verify all invariants in the table above. Most common culprit: missing center_crop=True when trained with image_aug=True.

Issue: LIBERO eval crashes with EOFError asking for dataset path

Fix: set LIBERO_CONFIG_PATH and write a non-interactive config before headless eval.

Issue: ALOHA client ROS import fails with libffi symbol errors

Fix: conda install -c conda-forge libffi

Issue: flash-attn install fails

Fix: export TMPDIR and PIP_CACHE_DIR to the same filesystem, retry with --no-cache-dir.

Issue: EGL teardown logs show EGL_NOT_INITIALIZED

Fix: treat as teardown noise unless exit code is non-zero. Set EGL env vars:

export MUJOCO_GL=egl PYOPENGL_PLATFORM=egl
export CUDA_VISIBLE_DEVICES=0 MUJOCO_EGL_DEVICE_ID=0

For HPC/cluster users

On Slurm clusters, route caches to scratch to avoid filling /home quota:

export HF_HOME=/scratch/$USER/.cache/huggingface
export XDG_CACHE_HOME=/scratch/$USER/.cache
export PIP_CACHE_DIR=/scratch/$USER/.cache/pip
export TMPDIR=/scratch/$USER/tmp

Avoid stacking cluster Python modules when using conda. Typically module load cuda is sufficient.

Advanced topics

Paper summary and checkpoints: See references/paper-and-checkpoints.md Detailed LIBERO workflow: See references/libero-workflow.md Detailed ALOHA workflow: See references/aloha-workflow.md Config map and troubleshooting matrix: See references/config-troubleshooting.md

Resources

• Project website: https://openvla-oft.github.io/
• Paper: https://arxiv.org/abs/2502.19645
• Repository: https://github.com/moojink/openvla-oft
• RLDS builder: https://github.com/moojink/rlds_dataset_builder