Remote Teleoperation & Physical Safety with the SO-101 Robot Arm

The Mission: Physical Tap Manipulation from Anywhere

The objective of this project is straightforward but physically high-consequence: from anywhere in the world, command an SO-101 6-DOF robotic arm over a secure private mesh (Tailscale) to open a kitchen water tap, fill a reservoir, close the tap, and verify completion on live video.

In software, an unhandled exception causes a crash; in plumbing robotics, an unhandled exception floods your apartment. Therefore, the entire architecture is driven by physical safety guarantees.

Non-Negotiable Safety Architecture

1. Passive Physical Overflow: The water reservoir is physically aligned such that any overflow spills directly into the sink basin. Even in a catastrophic total power loss while the valve is open, the failure mode is water draining safely down the pipe.
2. On-Board Monotonic Timer Deadline: The command to open the tap always specifies a finite duration (e.g. default_open_seconds = 30). The Raspberry Pi enforces this deadline via an independent hardware monotonic timer. Even if the laptop disconnects or network drops, the Pi closes the tap when the timer expires.
3. Emergency Action is Active Close (Not Torque-Off): In robotics, disabling servo torque is the usual emergency stop. In plumbing, torque-off leaves the tap handle open. The emergency handler commands an immediate move to the verified CLOSED pose at max torque.
4. Verified Position Gate: After any close command, the Pi reads back high-resolution bus telemetry (angle, temperature, current). If position is outside tolerance, it retries three times before latching into an unrecoverable FAULT state.

Dual Teleoperation & Teaching Modes

• Bilateral Leader-Arm Teleoperation: Using a matched 3D-printed SO-101 leader arm equipped with low-resistance potentiometers on my Mac workstation, I can physically manipulate the leader arm to intuitively record smooth end-effector trajectories.
• Web UI Teach Mode: A browser-based 3D jog controller streaming joint angle targets at ~30 Hz over WebSockets (/ws/jog), allowing remote fine-tuning of closed, pre-turn, and open poses.
• LeRobot Integration: Paving the way for autonomous policy rollout using Hugging Face’s LeRobot imitation learning framework (ACT / Diffusion Policy).