SENBAOZHENG/openvla-lora-oft-libero

OpenVLA-7B LoRA+OFT for LIBERO-10

This repository contains the best LoRA+OFT adapter produced by the OpenVLA Efficiency Project for continuous robot-action prediction on LIBERO-10.

It is an adapter-only release based on openvla/openvla-7b. It does not contain the full OpenVLA-7B base-model weights.

Model details

Item	Value
Base model	openvla/openvla-7b
Benchmark	LIBERO-10
Training dataset	openvla/modified_libero_rlds
Fine-tuning method	LoRA + continuous-action OFT head
LoRA rank	32
LoRA alpha	16
LoRA dropout	0.0
Batch size	12
Training steps	100,000
Random seed	42
Training objective	Continuous L1 action regression
Action representation	8-step continuous action chunks
Additional inputs	Proprioceptive state and wrist-camera observations

Evaluation

The author-reported closed-loop result under the project evaluation configuration is:

Metric	Result
LIBERO-10 closed-loop success rate	94.0%
Open-loop average inference latency	approximately 125 ms
Peak inference VRAM	approximately 14.4 GB
Inference precision	BF16

These results were obtained under the project's own simulation, hardware, preprocessing, and control configuration. They should not be interpreted as official OpenVLA benchmark results.

Repository contents

File	Purpose
adapter_model.safetensors	LoRA adapter weights
adapter_config.json	PEFT LoRA configuration
action_head.pt	Continuous OFT action-head weights
proprio_projector.pt	Proprioceptive-state projection weights
dataset_statistics.json	Action and state normalization statistics
dataset_statistics_provenance.json	Statistics provenance information
validation_metrics.json	Saved validation results
processor/	Tokenizer and image-processor configuration

Important loading requirement

This repository is not a standalone text-generation model.

A plain PEFT loader is not sufficient to reconstruct the complete robot policy because the release also contains a custom continuous action head and proprioception projector.

A complete loader must:

1. Load openvla/openvla-7b.
2. Attach adapter_model.safetensors using adapter_config.json.
3. Restore action_head.pt.
4. Restore proprio_projector.pt.
5. Load the supplied processor files.
6. Apply the supplied dataset normalization statistics.

The corresponding training, evaluation, and policy-loading code is available in the project repository:

https://github.com/SILVIO-ZHENG/openvla_efficiency_project

Download

hf download SENBAOZHENG/openvla-lora-oft-libero \
  --local-dir ./openvla-lora-oft-libero

Intended use

This model is intended for research involving:

• Vision-language-action policies
• Parameter-efficient OpenVLA fine-tuning
• Continuous robot-action prediction
• LIBERO simulation
• Action-chunk execution and rolling replanning

Limitations

• The reported evaluation was conducted in LIBERO simulation.
• The model has not been validated for safety-critical operation.
• Real-robot deployment requires robot-specific action conversion, calibration, timing control, and safety checks.
• Performance may change with different observations, normalization statistics, simulator versions, or control configurations.

References

• OpenVLA: An Open-Source Vision-Language-Action Model
• LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning

License

The adapter artifacts in this repository are released under Apache-2.0. Users must also comply with the licenses and usage conditions of the OpenVLA base model and the LIBERO training data.