SII-LiuLab / manimux

### Real robot experiments. One inference harness.
An open-source harness for real-world manipulation inference and experiment management.
**Website** · **Documentation** · **Quick start** · **Demo**
**RoboGUI** brings camera views, robot state, action chunks, and recorded trajectories
into one workspace.
Watch the real-robot demo · Experiment workflow
## What you can do
- **Run policies on real robots.** Deploy supported model and station configurations,
 inspect live observations, and control runs through RoboGUI.
- **Study inference strategies.** Compare supported scheduling and sampling methods
 within a shared experiment workflow.
- **Keep experiments traceable.** Save run configurations, trajectories, recordings,
 and task-defined human evaluations for review and comparison.
## Quick start
Try RoboGUI without hardware. Requires **Python 3.11 or 3.12** and uv.
 [code] 
Open **http://127.0.0.1:8086**. The preview uses synthetic data and does not connect
to a robot. Model environments and checkpoints are needed only for real deployment.
**Ready for hardware?** Follow a supported deployment recipe
and the station setup guide.
## Included integrations
| Area | Examples |
| --- | --- |
| Embodiments | YAM, Tianji–TacCap; experimental ARX X5 and PiPER adapters |
| Policies | Pi05, diffusion policies, GR00T, LingBot-VLA2, XR-1, and more |
| Inference | Serial and async chunking, RTC, ACT temporal ensembling, PAINT, and more |
These are individual integrations, not a fully compatible cross-product. Support
varies by checkpoint, backend, adapter, and station; experimental hardware adapters
have not completed physical validation.
Full support catalog and deployment recipes · Supported inference combinations
## Architecture
ManiMux sits between policy servers and robot hardware, coordinating observation/action
adaptation, scheduling, execution, and experiment records. Model frameworks own inference;
RoboGUI provides the operator workspace.
 [code] 
Architecture and extension protocols · Inference guide
## Contributing
Contributions to robot and policy integrations, inference strategies, RoboGUI,
and documentation are welcome. Start with the contribution guide.
For agent-assisted development, use the development skill.
**Latest update · 2026-10-18:** The ManiMux website and dedicated user guide are live.
## Dream-1 hosted inference
Run Dream-1 beta on a configured dual-arm YAM station with three cameras. No local
model server, checkpoint download or inference GPU is required. The
standalone example
uses a simple **30 Hz serial baseline**, executing all 48 returned actions before
requesting the next chunk. It does not use Smooth, blending or latency cropping;
API latency therefore produces a pause between chunks. This is a starting example,
not a claim of validated task performance.
First complete the station setup,
including the YAM SDK, camera bindings and your private
`manimux/configs/local/station.yaml`. From the repository root, use the Python
envir