SO-101 teleoperation: leader arms, 30 Hz recording and scaling past one desk

The SO-101 is an open-source 3D-printed arm co-designed by TheRobotStudio and Hugging Face, the successor to the SO-100 with improved wiring, easier assembly and new leader-arm motors. LeRobot's docs call it "our flagship robot", it is built into LeRobot as `so101_follower` and `so101_leader`. Its parts list comes to a few hundred dollars for a leader and follower pair. That low price is exactly why teams outgrow their first setup quickly: the arm is never the bottleneck. Operators, resets and consistency are.

How the SO-101 leader/follower pair works

Each arm has six Feetech STS3215 servos on a serial bus, connected to the computer over USB through a bus servo adapter. The follower uses six identical 1/345 motors. The leader mixes three gear ratios, 1/345 at the shoulder lift, 1/191 at the base and elbow, 1/147 at the wrist and gripper, so it can hold its own weight while staying light to move. The operator moves the leader by its handle and trigger; LeRobot reads the leader's joint positions and sends them to the follower as targets. There is no inverse kinematics anywhere in that loop, which is why leader-arm demonstrations come out smooth and natural.

Calibration is what makes this work across units: you move each arm to the middle of its ranges and then sweep every joint, so that the leader and the follower report the same values in the same pose. LeRobot's docs put it plainly: calibration "allows a neural network trained on one robot to work on another". On a fleet of arms, sloppy calibration is the most common reason a policy trained on one pair misbehaves on the next.

Other input devices

Control rate and latency for remote SO-101 operators

LeRobot's recording tutorial runs the SO-101 loop at 30 Hz, with cameras at 640×480 and 30 fps in its examples: read the observation, read the leader, send the action, sleep until the next frame. At 30 Hz each step is about 33 ms. A remote operator on a 120 ms round trip sees the result of a move four steps late. Pick-and-place on an SO-101 tolerates that; fast, contact-heavy motions do not.

Remote operation is where the SO-101's simplicity becomes a constraint. Both arms are USB devices, so the built-in loop assumes the leader sits next to the follower. For an operator in another city you need a bridge: the leader at the operator, its joint stream sent over the network to the robot host, and the follower loop kept local. LeRobot's third-party list includes a LiveKit (WebRTC) teleoperator plugin built for exactly this kind of remote control; a 3D mouse or phone sending deltas is the lower-effort alternative. Whatever you choose, keep the robot host on the robot's network and avoid remote desktop. Our latency guide covers budgets by task.

Recording SO-101 data that trains

`lerobot-record` writes a LeRobotDataset (Parquet plus video) and pushes it to the Hugging Face Hub unless you pass `--dataset.push_to_hub=False`. The defaults are 50 episodes, 60 s per episode and 60 s of reset time. During a session, the right arrow (or n) ends an episode early, the left arrow (or r) throws it away and re-records, and Escape (or q) stops, encodes and uploads. Those keys also work from a terminal over SSH, which matters when the operator is not in the room.

We deliver SO-101 episodes as LeRobot datasets by default, or as RLDS or HDF5 if your training stack needs them. The data collection guide has the full checklist.

Running several pairs at once

The cheapest way to more episodes per day is more SO-101 pairs, and the parts cost makes that practical. A few habits keep a multi-pair setup from turning into a mess:

Human-in-the-loop corrections on an SO-101

Once a first policy works, the most useful data is often corrections rather than fresh demonstrations. LeRobot's HIL data collection mode (`lerobot-rollout --strategy.type=dagger`) runs the policy, lets the operator pause, take over to recover and correct, then hand back, all inside one recorded episode. It needs a teleoperator with active motors that can move to mirror the robot when paused, and the SO-101 leader (`so_leader`, or `bi_so_leader` for two arms) is on the compatible list. That pause-and-mirror step avoids the follower jumping to wherever the leader was left. For VLA models such as smolVLA, Pi0 and Pi0.5 the same mode runs with real-time chunking.

Common SO-101 failure modes

SO-101, Koch v1.1, or a step up

The Koch v1.1 is the other leader/follower arm built into LeRobot, on a different servo family. The SO-101 wins on parts cost and on how much of LeRobot's documentation, from calibration to smolVLA and HIL collection, is written around it. Neither publishes payload or repeatability. When a task needs a heavier object, a longer reach or a documented precision spec, the usual next step is an arm such as the Trossen WidowX AI, which keeps the same LeRobot record-and-train loop through a maker plugin and publishes 1.5 kg payload and 1 mm repeatability. The stack you learned on the SO-101 transfers; the operator protocol transfers too.

How RoboPilots runs SO-101 data collection

Your arms stay with you. We agree the input path (leader arms with a network bridge, or a 3D mouse), set up the robot host on your side of a VPN, and train a dedicated group of operators on your units and your task protocol before the first recorded episode.