Policy + SiMa: compile robot policies (ACT, NVIDIA GR00T) for the SiMa.ai Modalix DevKit and drive a robot arm with them.
checkpoint ──compile──▶ bundle ──deploy──▶ DevKit ──robot run──▶ SO-101 arm
(your PC) (folder) (Modalix)
PoLiMa is the robot counterpart of LLiMa. It does not train models: train with LeRobot or NVIDIA's GR00T code as usual, and PoLiMa starts from the checkpoint.
| Model | Compile on Linux | Run on the DevKit | Tested on an SO-101 arm |
|---|---|---|---|
| ACT (LeRobot) | yes | yes | yes |
| GR00T N1.6 | yes | yes | yes (mixed precision) |
| GR00T N1.7 | yes | yes | yes (mixed precision) |
- Set up an SO-101. Equipment and deployment workflow: examples/so101.
- Run your own model. Record, train, compile, check and run: docs/your-own-model.md.
Start with installation, including host prerequisites,
DevKit installation and access from a laptop to a remote compiler host.
After installation, run polima workbench start and open its printed URL.
Download the SO-101 pick-and-place bundles from Hugging Face: GR00T N1.6 and GR00T N1.7. Both produce 32 actions per plan. Follow the download and setup instructions. If repository access is restricted, sign in with an authorized Hugging Face account.
The SO-101 pick-and-place setup. Follow the hardware, measurements and camera-view guide to reproduce the arrangement.
Workbench runs on the build host: import a checkpoint and dataset, configure compilation, and deploy the resulting bundle to a DevKit.
Studio runs on the DevKit: select a policy, configure hardware, manage the runtime and benchmark inference before running the arm.
These navigation previews use example data; they do not show a hardware run or measured performance. See the SO-101 setup guide for the reference hardware, dataset, camera placement and calibration instructions.
On your computer:
| Command | What it does |
|---|---|
polima compile |
Turns a checkpoint into a bundle: the compiled model, its settings and the robot client. |
polima deploy |
Copies a bundle to the DevKit and makes it the active one. |
polima run --fixture |
Checks a deployed bundle's answers and speed, without the arm. |
polima workbench start |
Web page for your checkpoints, bundles and DevKits (README). |
polima doctor |
Checks this computer, the compiler and the DevKit. |
On the DevKit:
| Command | What it does |
|---|---|
polima robot calibrate |
Calibrates the arm once. |
polima robot run |
Drives the arm with the active bundle. Ctrl-C stops and returns the arm to rest. |
polima studio start |
Web page to install bundles, watch the cameras and run the arm (README). |
polima server start / stop / status |
The model server. |
Every command has --help.
Reference measurements on a Modalix DevKit, mixed precision, one 32-move plan:
| Model | Model time on the DevKit | Camera picture to plan |
|---|---|---|
| GR00T N1.6 | 317 ms | 364 ms |
| GR00T N1.7 | 233 ms | 285 ms |
These measurements describe the reference setup, not a guaranteed latency for every checkpoint or release bundle. Use the model card for artifact-specific validation and benchmark your deployed bundle in Studio.
The arm keeps moving while the next plan is computed.
| Folder | Contents |
|---|---|
src/polima/ |
The polima command: compile, deploy, robot client, Studio, Workbench. |
runtime/ |
The C++ model server that runs on the DevKit (README). |
examples/so101/ |
Our SO-101 setup, and arm and teleoperation tools. |
docs/ |
Reference guides: install, your own model, GR00T, deploy, robot. |
PoLiMa's code is licensed under Apache-2.0. Bundled UI component notices are included in LICENSE. GR00T bundles remain subject to NVIDIA’s model licence, including the non-commercial research/evaluation restriction in the source checkpoint licences. See third-party licences and the licence supplied with each model package.



