README.md

:new: Updates

• [07/2026] :fire: We propose BehaviorVLA, a follow-up to OptimusVLA on memory-driven VLA models.
• [05/2026] :fire: We release the Inference Code and Checkpoints on LIBERO.
• [02/2026] :fire: OptimusVLA is accepted to CVPR 2026!
• [02/2026] :fire: Project page released.
• [02/2026] :fire: Arxiv paper released.

:balloon: OptimusVLA Framework

Overview of OptimusVLA framework. Given a task and the current observation, the Vision–Language backbone first encodes the inputs into a multimodal representation. GPM then retrieves a task-level prior based on this representation, while LBM dynamically encodes the historical action sequence to produce a consistency constraint. Finally, the flow policy denoises the initialization with an adaptive NFEs schedule to generate the action chunk.

:rocket: How to Run

OptimusVLA is built on the openpi framework. Therefore, please first download and configure the openpi environment, and then download the pi_05 model weights.

Install openpi

1. Clone the official OpenPI repository

git clone --recurse-submodules git@github.com:Physical-Intelligence/openpi.git
cd openpi

2. Create the main OpenPI environment:

GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

3. PyTorch Support

cp -r ./src/openpi/models_pytorch/transformers_replace/* .venv/lib/python3.11/site-packages/transformers/

4. Download pi05_libero model checkpoints and convert it to pytorch version

OptimusVLA does not release the pi0.5 policy checkpoint. Prepare a PyTorch pi05_libero checkpoint yourself. The policy directory must contain:

model.safetensors
assets/physical-intelligence/libero/norm_stats.json

If you start from a JAX OpenPI checkpoint, convert it with the upstream OpenPI converter:

cd "${OPENPI_ROOT}"
uv run examples/convert_jax_model_to_pytorch.py \
  --checkpoint-dir /path/to/pi05_libero_jax_checkpoint \
  --config-name pi05_libero \
  --output-path /path/to/pi05_libero_pytorch

Set the policy path:

export POLICY_DIR=/path/to/pi05_libero_pytorch

5. Install the extra OptimusVLA inference dependency:

uv pip install faiss-cpu
uv pip install mamba-ssm

6. Create the LIBERO Client Environment

Create the LIBERO example environment using the official OpenPI instructions:

cd "${OPENPI_ROOT}"
uv venv --python 3.8 examples/libero/.venv
source examples/libero/.venv/bin/activate
uv pip sync examples/libero/requirements.txt third_party/libero/requirements.txt \
  --extra-index-url https://download.pytorch.org/whl/cu113 \
  --index-strategy=unsafe-best-match
uv pip install -e packages/openpi-client
uv pip install -e third_party/libero
deactivate

Use this environment only for the LIBERO client in examples/libero/main.py. Run the policy server from the OpenPI server environment created in step 2.

Apply the OptimusVLA Code

Clone or download the GitHub repository that contains this code/ folder, then copy the overlay files from code/ into an already configured OpenPI checkout:

git clone https://github.com/iLearn-Lab/CVPR26-OptimusVLA.git
cd CVPR26-OptimusVLA
export OPENPI_ROOT=/path/to/openpi
rsync -av \
  code/ "${OPENPI_ROOT}/"

This step only copies the source overlay. Download checkpoints/ and memory/ separately from the Hugging Face asset repository in the next step.

Download OptimusVLA Assets

Download the asset directories from the OptimusVLA_Memory. They must end up under the OpenPI root exactly as follows:

${OPENPI_ROOT}/checkpoints/gpm_task_head.pt
${OPENPI_ROOT}/checkpoints/lcm.pt
${OPENPI_ROOT}/memory/gpm_memory_meta.pt
${OPENPI_ROOT}/memory/gpm_memory.index
${OPENPI_ROOT}/memory/gpm_memory_actions.npz

Run LIBERO Evaluation

cd "${OPENPI_ROOT}"
POLICY_DIR=/path/to/pi05_libero_pytorch bash scripts/run_libero_eval.sh

By default, the script runs libero_spatial, libero_object, libero_goal, and libero_10. Logs are written to a timestamped directory under logs/, and the final summary is saved as results.txt in that directory. The summary contains the exit code, episode count, success count, and success rate for each suite.

Default model and server settings:

Parameter	Default	Meaning
POLICY_CONFIG	pi05_libero	OpenPI policy config used for the base pi0.5 checkpoint.
POLICY_DIR	Required	Local PyTorch pi0.5 checkpoint directory. This release does not include it.
ACTION_NORM_STATS_PATH	${POLICY_DIR}/assets/physical-intelligence/libero/norm_stats.json	Action normalization stats used by GPM memory actions.
HOST / PORT	127.0.0.1 / 8000	Client connection target and server port.
SERVER_CUDA_VISIBLE_DEVICES	${CUDA_VISIBLE_DEVICES} or 0	GPU used by the policy server.
CLIENT_CUDA_VISIBLE_DEVICES	${CUDA_VISIBLE_DEVICES} or 0	GPU visible to the LIBERO clients.
OPENPI_TORCH_COMPILE	0	Disables Torch compile by default for easier first-run debugging.

Default GPM settings:

Parameter	Default	Meaning
GPM memory	Enabled	The helper always passes --use-memory.
GPM assets	checkpoints/gpm_task_head.pt, memory/gpm_memory_meta.pt, memory/gpm_memory.index, memory/gpm_memory_actions.npz	Default task head, metadata, FAISS index, and packed action memory paths.
--action-use-quantile-norm	Enabled	Uses LIBERO action quantile statistics for memory action normalization.
MEMORY_TOP_K	8	Number of retrieved memory candidates per query.
MEMORY_REFRESH_EVERY	1	Refreshes memory retrieval every replan request.
ALIGN_MODE	hybrid	Time alignment mode for the retrieved memory trajectory.
MIXTURE_MODE	gaussian	Builds a Gaussian action prior from retrieved memories.
TEMPERATURE	10.0	Softmax temperature used when weighting retrieved memories.
SIGMA_MIN	0.05	Lower bound for the Gaussian prior standard deviation.
NOISE_MIN / NOISE_MAX	0.20 / 1.00	Noise schedule range used by the memory-guided sampler.
NFE_MIN / NFE_MAX	1 / 10	Sampling step range selected from memory confidence.

Default LCM settings:

Parameter	Default	Meaning
USE_LCM	1	Enables LCM refinement after GPM. Set USE_LCM=0 for GPM-only inference.
LCM_SCALE	0.10	Strength of the LCM correction applied to the action chunk.
LCM checkpoint	checkpoints/lcm.pt	Default LCM checkpoint path.
LCM architecture fallback	hidden 256, layers 1, heads 4, dropout 0.0, mamba_impl=auto	Used only when these fields are absent from checkpoint metadata.

Default evaluation and logging settings:

Parameter	Default	Meaning
Suites	libero_spatial, libero_object, libero_goal, libero_10	Four standard LIBERO suites run in parallel.
NUM_TRIALS_PER_TASK	50	Episodes evaluated per task.
REPLAN_STEPS	10	Number of actions executed before requesting a new chunk.
NUM_STEPS_WAIT	10	Initial dummy steps before policy control begins.
SEED	7	LIBERO environment seed.
RESIZE_SIZE	224	Image size sent to the policy client.
MUJOCO_GL	egl	Headless MuJoCo rendering backend.
LOG_DIR	logs/libero_eval_<timestamp>	Directory for server logs, client stdout logs, JSONL records, and results.txt.
RESULTS_TXT	${LOG_DIR}/results.txt	Final text summary with per-suite and overall success rates.

Experimental Results Note

Our repeated experiments show that performance on LIBERO is affected by the GPU and randomness. Below, we provide the experimental results obtained on L40 and A800 GPUs using the default parameter settings for reference.

Method	Spatial	Object	Goal	Long	Average
pi_05	98.8	98.2	98.0	92.4	96.9
OptimusVLA (L40)	99.4	99.2	98.8	95.6	98.3
	98.6	99.8	97.8	94.6	97.7
	97.8	99.6	98.6	92.8	97.2
	98.4	99.0	98.0	95.2	97.7
	98.6	99.4	98.0	94.6	97.7
	98.8	98.8	97.2	95.2	97.5
	98.4	99.8	97.4	94.0	97.4
	98.4	99.4	97.6	95.0	97.6
OptimusVLA (A800)	97.6	99.4	96.8	95.6	97.4
	97.8	99.4	98.2	95.2	97.7
	99.6	99.2	99.0	94.4	98.1
	98.4	99.6	97.8	94.8	97.7
	99.0	99.6	97.0	94.2	97.5
	99.2	99.2	97.4	94.0	97.5
	98.6	99.8	98.8	96.4	98.4
	99.2	99.2	98.4	94.0	97.7
OptimusVLA (Paper)	99.6	99.8	98.4	96.4	98.6

:smile_cat: Evaluation results on Real World

We evaluate OptimusVLA on Generalization Tasks and Long-horizon Tasks via GALAXEA R1 Lite robot.

:hugs: Citation

If you find this work useful for your research, please kindly cite our paper:

@article{li2026optimusvla,
  title={Global Prior Meets Local Consistency: Dual-Memory Augmented Vision-Language-Action Model for Efficient Robotic Manipulation},
  author={Zaijing Li, Bing Hu, Rui Shao, Gongwei Chen, Dongmei Jiang, Pengwei Xie, Jianye Hao, Liqiang Nie},
  journal={arXiv preprint arXiv:2602.20200},
  year={2026}
}