Pi0.5 SO-101 Pick-and-Place
A task-specific Pi0.5 policy fine-tuned to place a grey eraser in a white basket, compiled for the SiMa.ai Modalix DevKit.
| Property | Value |
|---|---|
| Base model | LeRobot Pi0.5 |
| Training checkpoint | Step 20,000 |
| Compiled output | 50 actions × 6 joints |
| Denoising steps | 10 |
| Precision | W8A16; shared attention in BF16 |
| Compiled graphs | 38 |
| Robot | SO-101 follower |
| Task | Place the grey eraser in the white basket |
This repository contains the deployable compiled bundle, reference fixtures and SO-101 runtime assets. It does not contain the training checkpoint or dataset. The policy is specific to the reference task and workspace.
Prerequisites
- Modalix DevKit with NVMe storage and a compatible SiMa.ai runtime.
- A matching PoLiMa source checkout on the deployment host and DevKit.
- A Linux deployment host, or WSL on Windows, with SSH, rsync and the Hugging Face CLI.
- SO-101 follower with six STS3215 motors (IDs 1–6), securely mounted and powered.
- Logitech C920 overhead camera and a USB wrist camera, connected to the DevKit and framed like the training setup.
The release checks used ModelSDK 2.1.3 and LLiMa 0.4.0.
Download and deploy
hf download simaai/Pi0.5-SO101-PickPlace --local-dir Pi0.5-SO101-PickPlace
cd Pi0.5-SO101-PickPlace
sha256sum --check SHA256SUMS
read -r -p "DevKit login (username@address): " devkit_login
polima deploy --bundle ./bundle --devkit "$devkit_login" --start
Deployment installs and activates the bundle and starts the inference server. It does not start robot motion.
Check inference
polima run --bundle ./bundle --devkit "$devkit_login" --fixture --repeat 20
This replays the packaged observations without cameras or robot motion and checks numerical agreement. It does not measure physical task success.
Run the robot
Use Studio to inspect both camera views and confirm the selected arm, cameras and task. Calibrate each physical follower arm before its first run.
polima studio start
polima robot calibrate
polima robot run
The reference cameras capture 640 × 480 MJPG at 30 frames/s. The overhead lens was 49 cm above the table; the basket was approximately 32 cm from the arm end. The basket measured 17.5 × 12 × 6 cm and the eraser 12.5 × 5 × 2 cm.
Training data
Fine-tuned on 769 demonstrations from SO-101 Grey Eraser Pick-and-Place. The full collection contains 961 episodes and 272,853 frames.
Validation
On the reference Modalix runtime, 20 measured requests after three warm-ups gave 358.5 ms p50, 359.1 ms p95 and 360.3 ms p99 model-server latency. The combined vision-language prefill took 202.39 ms and ten denoising steps took 155.26 ms. The packaged eight-case fixture check passed the W8A16 fidelity envelope against the PyTorch checkpoint. Action mean absolute error was 0.016 overall and 0.037 at the worst timestep. Physical operation was verified on the reference SO-101 setup; a controlled success rate was not recorded.
Package contents
bundle/ contains 38 compiled graphs: one vision graph, 18 prefix input graphs,
one shared attention graph, 17 prefix output graphs and one action-expert graph.
It also contains the execution plan, runtime constants, tokenizer, embedding
table, normalization statistics, reference observations and SO-101 launch
assets. SHA256SUMS covers every supplied file except the checksum file itself.
Limitations
The policy is specific to the reference robot, object, basket and camera views. Different geometry, framing, lighting or calibration can change behavior. A new text instruction can select behavior learned by the fine-tuned checkpoint; it does not teach a new task at runtime.
License
This model is derived from lerobot/pi05_base, whose model card specifies the
Gemma license. Review those terms
before use or redistribution. PoLiMa's source code has its own Apache-2.0 terms.