cix-vla-app is a browser-based robot control console built on top of lerobot, designed as a control platform for CIX VLA deployments based on the CIX P1 SoC. It packages runtime control, model serving, voice interaction, automation, telemetry, and operator-facing documentation into one application for target-device operation, validation, and deployment.
Runtime model files are not fully included in this repository. Download them from ModelScope: cix/cix-vla-app, then merge the corresponding directories into this project (for example, under model_zoo/ and omni/llm/).
CIX-VLA APP is primarily developed for the CIX P1 platform. In this project, the CIX P1 serves as the target compute platform for browser-based runtime control, model loading, voice interaction, automation orchestration, and operator-side task execution.
Based on publicly available platform information for CIX P1-based boards, the CIX P1 highlights include:
- 6nm process technology
- 12-core Arm CPU with a tri-cluster design
- up to 45 TOPS heterogeneous AI computing capability
- Arm Immortalis GPU class graphics support on public board platforms
- high-speed peripheral support such as PCIe Gen4 and modern memory/storage options on target boards
From an application perspective, the project uses the CIX P1 as a local VLA deployment platform where heterogeneous compute, media, and I/O capabilities can be combined in one target system for:
- edge-side runtime control and model serving
- local audio, camera, and device access
- task execution and automation orchestration on the target system
- operator-facing debugging and validation workflows for deployed VLA capabilities
This makes the platform especially practical for scenarios that need local perception, task execution, model inference, and operator interaction to run close to the device.
Note: exact exposed interfaces and final specifications may vary by carrier board or system product built around the CIX P1.
- Runtime control for powering the system on or off, loading or unloading models, and projecting runtime state to the UI
- Manual task execution for the supported tabletop tasks such as tape and pen operations
- Voice-command interaction with browser-side audio capture and Omni-based task classification
- Detection and workflow orchestration controls for semi-automatic and automatic operation flows
- Bilingual operator documentation, help pages, blogs, and debug views
- Deployment helper scripts for virtual cameras, udev aliases, and performance collection
The following demos show representative operator flows supported by the current release.
The animation below shows the current model-loading dialog flow in the master line.
The animation below shows the manual-task interaction flow. Operators can click a supported task button and the VLA runtime will execute the corresponding task on the target device.
The animation below shows the automation flow. Operators can configure the task order in advance, and the VLA runtime will execute the arranged workflow automatically.
The animation below shows a voice-command demo where the recognized instruction is to retrieve the tape from the box, and the VLA runtime executes the corresponding task.
The animation below shows a voice-command demo where the recognized instruction is to place the tape back into the black box, and the VLA runtime executes the corresponding task.
Current voice-command demos are configured for Chinese instructions for the supported task set. If you want to enable this capability in your own environment, use the bundled Omni model assets under
omni/llm/and follow the upstream Qwen Research License terms. SeeTHIRD_PARTY_LICENSES.md. If you have different task wording or language requirements, you can customize the Omni prompt configuration inconfig/audio.yamlunderomni_prompt_matchto adapt the recognition behavior.
app.pyApplication entry point for the Gradio-based operator consolesrc/ui/Pages, components, event wiring, runtime polling, and bilingual UI behaviorsrc/runtime/Runtime configuration loading, runner logic, patch management, and process coordinationsrc/services/Voice, orchestration, Redis state, and UI-facing status servicessrc/telemetry/Camera, action, and telemetry collection helpersmodel_zoo/Model assets, overlay payloads, and compatibility supportomni/Omni runtime resources used by the voice command flowscripts/Operational helper scripts such as virtual-camera setup and performance capture
This project is intended for:
- operator-side control of a CIX VLA robot arm
- validation and debugging of runtime, voice, and automation behavior
- deployment on CIX P1-based target devices with local cameras, audio input, and model assets prepared
The first-party code in this project is licensed under Apache-2.0. Third-party components keep their own licenses. See LICENSE and THIRD_PARTY_LICENSES.md in this directory for the final licensing boundary.