hello-hands

An SO-101 leader–follower robot arm running LeRobot, built toward collecting my own demonstration data and fine-tuning Physical Intelligence's π0 model via openpi.

This repository is code + docs only. No training data, video, or model checkpoints live here — those are published to the Hugging Face Hub (see Datasets & models).

Status

• ✅ Both arms assembled and calibrated — leader and follower, 6 joints each (Feetech STS3215 servos). Calibration is committed under calibration/.
• ✅ Teleoperation works on all 6 joints. Along the way the base joint (shoulder_pan) had an encoder wraparound bug — diagnosed and fixed; see the base-joint encoder bug.
• ✅ Both cameras installed — wrist (InnoMaker 32×32 UVC, on the follower gripper) and scene (Logitech C920s), 640×480 @ 30 fps.
• ✅ 50 demonstrations recorded and published — 22,500 frames, two camera views — dataset on the Hub.
• ✅ SmolVLA fine-tuned locally on that dataset and deployed autonomously via lerobot-rollout (leader disconnected, RTC inference).
• 🔧 The policy approaches the correct object but fails the grasp — the task is not yet completed autonomously. Dataset #2 (rigid object, denser start region) is the next pass.
• ⬜ π0 fine-tuning via openpi — not started.

A fine-tuned policy now exists and pursues the right object, but the task is not yet completed autonomously and π0 is untouched; this README claims only what is actually complete.

The pipeline

• Assemble the leader and follower arms
• Calibrate every joint (per-motor range + homing offset)
• Teleoperate — leader drives follower across all 6 joints
• Install cameras
• Collect demonstrations — teleoperate while recording camera + joint data into a LeRobotDataset
• Validate the full loop locally — fine-tune SmolVLA and deploy it autonomously (approaches the object; grasp not yet reliable)
• Fine-tune π0 with LoRA on a rented GPU (via openpi) — imitation learning (behavior cloning), not RL
• Deploy the trained policy behind an openpi policy server
• Iterate on data quality — more/better demonstrations, retrain — dataset #2 (rigid object, denser start region) is the planned next pass

Exact commands for the completed and in-progress steps are in scripts/commands.md.

Hardware

Role	Arm type	Port	LeRobot id
Follower	so101_follower	COM4	my_follower
Leader	so101_leader	COM5	my_leader

• 6× Feetech STS3215 servos per arm (12-bit absolute encoders).
• Wrist camera: InnoMaker 32×32 UVC on a printed plug mount (follower gripper), 640×480 @ 30 fps, MJPG forced.
• Scene camera: Logitech C920s on a desk clamp, 640×480 @ 30 fps, MJPG forced.

The base-joint encoder bug

The follower's base joint (shoulder_pan) would drive into a hard stop under torque. Root cause: its encoder zero landed on the 0 / 4096 wraparound seam of the 12-bit absolute encoder, which pinned homing_offset at its rail (±2047) and pushed the joint's forward target outside its usable swept range.

Fix, without disassembly: hold the arm at forward-center and issue the Feetech "one-key middle" command (write 128 to the Torque_Enable register) so the servo re-homes its center to 2048, then recalibrate.

→ scripts/fix_base_encoder.py — small, heavily-commented, runnable. The header documents the bug in full.

The committed follower calibration is the post-fix state — its shoulder_pan homing_offset (-2014) sits off the ±2047 rail — while the leader's railed base offset (-2047) is a benign torque-off artifact, since the leader is never driven under power.

Repository layout

hello-hands/
├── .gitignore
├── README.md
├── LICENSE                     # MIT
├── scripts/
│   ├── camera_stress_test.py   # standalone camera/USB diagnostic
│   ├── fix_base_encoder.py     # base-joint encoder one-key-middle fix
│   └── commands.md             # LeRobot CLI reference for this hardware
└── calibration/
    ├── my_follower.json        # follower calibration (reproducibility evidence)
    └── my_leader.json          # leader calibration

(A docs/ folder with longer build write-ups is planned but not yet written.)

Datasets & models

Published to the Hugging Face Hub as the project reaches each step. The dataset is live; the π0 policy does not exist yet:

• Demonstration dataset — harrison-powe/hello-hands-pick-place_20260721_163414 — 50 episodes / 22,500 frames, two camera views
• Fine-tuned π0 policy — TODO: https://huggingface.co/<hf-username>/<model-name>

(The interim SmolVLA checkpoints are local-only and not published.)

Upstream

• TheRobotStudio/SO-ARM100 — arm hardware (SO-100 / SO-101)
• huggingface/lerobot — robotics library
• Physical-Intelligence/openpi — π0 model + serving

License

MIT © 2026 Harrison Powe