LingBot-VLA Runtime for RLinf / RPent

LingBot-VLA is a Vision-Language-Action model for general-purpose robot manipulation.

rlinf-lingbotvla is the RLinf-maintained inference runtime distribution used by RPent, including the LingBot-VLA model and deployment code required by the RoboTwin integration.

This distribution is focused on inference. It is not a replacement for the full upstream LingBot-VLA training environment.

Installation

Standalone inference runtime

pip install rlinf-lingbotvla==0.1.1

RPent

RPent users should normally install LingBot-VLA as part of the RoboTwin stack:

uv pip install -e ".[robotwin]"

RPent currently composes:

RPent
├── rlinf-robotwin-runtime
├── rlinf-lingbotvla
└── NVIDIA cuRobo

Runtime Contract

The 0.1.x inference runtime is validated with:

These versions describe the validated RLinf / RPent inference runtime.

They intentionally differ from the requirements documented for the complete upstream LingBot-VLA training environment.

Package Scope

The wheel contains the model and deployment packages required for inference:

deploy/
lingbotvla/

Repository-level training scripts, task utilities, tests, checkpoints, and datasets are not bundled into the wheel.

Model checkpoints must be obtained separately. See the upstream LingBot-VLA project for the published model weights and full model documentation.

Zero-LeRobot Inference Runtime

LeRobot is not required by the RPent inference path.

The RLinf runtime uses:

• transformers.PretrainedConfig for model configuration
• torch.nn.Module for the policy runtime
• lazy loading for training/data-only modules

As a result:

pip install rlinf-lingbotvla

does not install LeRobot.

The inference implementation was validated against the previous runtime before release.

Training

A train extra is provided for dependencies that are outside the inference path:

pip install "rlinf-lingbotvla[train]==0.1.1"

This extra contains supporting training/data dependencies maintained by this distribution.

Full LingBot-VLA training still requires the upstream training setup and its additional dependencies, including the upstream LeRobot-based data/training workflow.

For complete post-training instructions, dataset conversion, robot configuration, and checkpoint preparation, use the upstream documentation:

https://github.com/Robbyant/lingbot-vla

RPent Integration

This package is maintained as the VLA side of the RPent RoboTwin runtime:

RoboTwin environment
        ↓
rlinf-robotwin-runtime

LingBot-VLA inference
        ↓
rlinf-lingbotvla

motion planning
        ↓
NVIDIA cuRobo

        ↓
      RPent

Source repositories:

• RLinf runtime: https://github.com/RLinf/lingbot-vla
• Upstream LingBot-VLA: https://github.com/Robbyant/lingbot-vla
• RPent: https://github.com/RLinf/RPent

License and Attribution

LingBot-VLA is based on the upstream Robbyant LingBot-VLA project and retains its Apache-2.0 licensing and upstream attribution.

For the technical report, citation information, checkpoints, datasets, and full training documentation, please refer to the upstream project.