Original LeRobot Operation

Section: lerobot/ — a single SO-101 / SO-100 arm, LeKiwi Prerequisite: read through deployment/ first and get into the container The other path: xlerobot/ (the whole XLeRobot) Next: troubleshooting/

Use the lerobot-* command line to operate a single SO-101 / SO-100 arm or a LeKiwi.

⚠️ To run the whole XLeRobot, go back to xlerobot/. These command-line tools don't support xlerobot — for why, see section 0 below.

0. When to take this path

Your hardware	Which path	Why
Single SO-101 / SO-100 arm	✅ This page (`lerobot-*` CLI)	The original tools support it directly
LeKiwi	✅ This page (`lerobot-*` CLI)	The original tools support it directly
Whole XLeRobot (dual arms + base)	❌ Go to 04	There's no `xlerobot` in the CLI whitelist

Why two paths: command-line tools like lerobot-calibrate hard-code a robot import list in the source (14 original robots — full list in section 11), and xlerobot isn't on it. Force it through and you get:

lerobot-calibrate: error: argument --robot.type: invalid choice: 'xlerobot'

Conversely, the scripts under examples/ are written for the whole XLeRobot and are of no use for a single arm.

The two paths share the same environment — the package installed in the image is just called lerobot, and XLeRobot merely adds three more pieces of code to the robots/ directory. So you can switch at any time without reinstalling anything, and both can be imported together in the same process.

1. The complete workflow

Setup          Calibration    Using it                Training       Deploy
─────          ───────────    ────────                ────────       ──────
find-port  →  calibrate  →  teleoperate  ┐
find-cameras              record        ┴→  train  →  rollout
                          replay

One command per step — each is covered in turn below.

2. Setup: find the serial port and cameras

bash

lerobot-find-port          # interactive: unplug and replug the servo cable as prompted, then it prints the port name
lerobot-find-cameras opencv   # try each camera in turn and save a screenshot

lerobot-find-cameras saves its screenshots to outputs/captured_images/; if you see a picture, the camera works. Verified working (it detected /dev/video0 and captured successfully).

Both commands are interactive — adding --help directly raises an EOF, which is normal.

3. First time with new servos: write IDs

Brand-new servos all ship with ID 1; you have to change them one by one to 1–6 before they can share one bus:

bash

lerobot-setup-motors --robot.type=so101_follower --robot.port=/dev/ttyACM0

It prompts you to "connect only one servo, then press Enter"; when one is done, plug in the next.

Only needed on a freshly assembled arm, or when you've swapped servos. Arms that already communicate fine can skip this step.

4. Calibration

bash

lerobot-calibrate \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.id=my_arm

No calibration file → guided flow: first move the arm to the middle of its range and press Enter, then walk each joint through its full range
Calibration file exists → asks whether to reuse it or recalibrate

The calibration file lives at ~/.cache/huggingface/lerobot/calibration/robots/so101_follower/my_arm.json. Container startup mounts -v ~/.cache/huggingface (see 03-Deployment), so restarting the container doesn't lose it.

For a leader-follower pair, calibrate the follower and the leader once each:

bash

lerobot-calibrate --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower
lerobot-calibrate --teleop.type=so101_leader --teleop.port=/dev/ttyACM1 --teleop.id=my_leader

5. Teleoperation

Move the follower with the leader arm (leader-follower mode)

bash

lerobot-teleoperate \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --teleop.type=so101_leader  --teleop.port=/dev/ttyACM1 --teleop.id=my_leader

No leader arm? Use the keyboard

--teleop.type includes keyboard, which needs no second piece of hardware:

bash

lerobot-teleoperate \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --teleop.type=keyboard

You can also use --teleop.type=gamepad with an Xbox controller (needs pip install "lerobot[gamepad]").

With cameras for visual feedback

bash

lerobot-teleoperate \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --teleop.type=so101_leader  --teleop.port=/dev/ttyACM1 --teleop.id=my_leader \
  --display_data=true --display_mode=rerun

Using rerun visualization likewise requires the two X11 libraries installed first — see xlerobot/, section 1.

6. Collect demonstration data

bash

lerobot-record \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --robot.cameras="{ front: {type: opencv, index_or_path: /dev/video0, width: 640, height: 480, fps: 30} }" \
  --teleop.type=so101_leader  --teleop.port=/dev/ttyACM1 --teleop.id=my_leader \
  --dataset.repo_id=<your-username>/<my-first-dataset> \
  --dataset.num_episodes=10 \
  --dataset.single_task="pick up the cube"

Parameter reference:

Parameter	Purpose
`--robot.cameras`	Camera configuration, nested-dict syntax: `name: {type: opencv, index_or_path: ..., width/height/fps}`
`--dataset.repo_id`	The dataset's ID on Hugging Face (a dataset recorded locally can be pushed up there too)
`--dataset.num_episodes`	How many episodes to record
`--dataset.single_task`	Task description, used at training time
`--dataset.root`	Add this to store somewhere else; defaults to under the cache directory

Keys during recording (the terminal prompts you):

Key	Action
`→` (right arrow)	End the current episode early and move on to the reset phase
`←` (left arrow)	Discard the current episode and record it again
`Esc`	Stop the whole recording session

As soon as recording is done you can use the data locally. To push it to the Hub:

bash

--dataset.push_to_hub=true

(Remember the container is started with -e HF_ENDPOINT=https://hf-mirror.com.)

7. Replay to check

After recording, replay it once to confirm the data is good before training:

bash

lerobot-replay \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --dataset.repo_id=<your-username>/<my-first-dataset> \
  --dataset.episode=0

8. Training

bash

lerobot-train \
  --policy.type=act \
  --dataset.repo_id=<your-username>/<my-first-dataset>

lerobot-train doesn't need robot.type (it only takes a dataset and a policy type), so this command is shared with the XLeRobot path — data you recorded with XLeRobot trains the same way.
Training is very compute-hungry. On CPU you can only run tiny scales; for anything practical you need a Jetson's GPU.

Optional policies (--policy.type): act (most commonly used), diffusion, smolvla, pi0 / pi05 / pi0_fast, groot, tdmpc, vqbet, and more.

9. Real-robot deployment: run a trained policy

bash

lerobot-rollout \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --policy.path=<your-username>/<my-model> \
  --strategy.type=base

Parameter	Description
`--policy.path`	The trained model (an ID on the Hub or a local path)
`--strategy.type`	`base` is most common; there are also `sentry`, `highlight`, `episodic`, `dagger`
`--inference.type`	`sync` (default) or `rtc`

⚠️ Don't use lerobot-eval on a real robot — it's for simulation benchmarks (--env.type=aloha/pusht/libero/metaworld...), not real-robot deployment. Real robots use lerobot-rollout.

10. Full table of command-line tools

Command	Purpose	Needs `robot.type`?
`lerobot-find-port`	Find the serial port	No
`lerobot-find-cameras`	Find cameras	No
`lerobot-setup-motors`	Write IDs to new servos	Yes
`lerobot-calibrate`	Calibrate	Yes
`lerobot-teleoperate`	Teleoperate	Yes
`lerobot-record`	Record a dataset	Yes
`lerobot-replay`	Replay data	Yes
`lerobot-train`	Train a policy	No (shared with XLeRobot)
`lerobot-rollout`	Deploy a policy on a real robot	Yes
`lerobot-eval`	Simulation benchmark evaluation	No (uses `--env.type`)
`lerobot-edit-dataset`	Edit a dataset	No
`lerobot-dataset-viz`	Dataset visualization	No
`lerobot-info`	Print environment and version info	No

11. Robot types the CLI supports (full list)

These are all the types the lerobot-* command line actually recognizes (read out of the whitelist in the source). There's no xlerobot — that's why the tutorial splits into two paths.

Robots (--robot.type) — 14 types:

openarm_follower      bi_openarm_follower     rebot_b601_follower
bi_rebot_b601_follower so100_follower         so101_follower
bi_so_follower        hope_jr_hand            hope_jr_arm
koch_follower         lekiwi                  lekiwi_client
omx_follower          unitree_g1

Teleoperators (--teleop.type) — including:

so101_leader      so100_leader      bi_so_leader      koch_leader
openarm_leader    omx_leader        rebot_102_leader  unitree_g1
keyboard          gamepad           homunculus_arm    homunculus_glove

For the definitive list, rely on --help in your own image — different versions add and remove types.

12. A complete example (from zero to running)

bash

# 1. Find the devices
lerobot-find-port                          # note the port name; /dev/ttyACM0 is used below
lerobot-find-cameras opencv                # note the camera; /dev/video0 is used below

# 2. Write IDs to new servos (only needed for a freshly assembled arm)
lerobot-setup-motors --robot.type=so101_follower --robot.port=/dev/ttyACM0

# 3. Calibrate
lerobot-calibrate --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower

# 4. Try the keyboard to see whether it moves
lerobot-teleoperate \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --teleop.type=keyboard

# 5. Record 10 episodes
lerobot-record \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --robot.cameras="{ front: {type: opencv, index_or_path: /dev/video0, width: 640, height: 480, fps: 30} }" \
  --teleop.type=keyboard \
  --dataset.repo_id=<your-username>/my-first-dataset \
  --dataset.num_episodes=10 --dataset.single_task="pick up the cube"

# 6. Train
lerobot-train --policy.type=act --dataset.repo_id=<your-username>/my-first-dataset

# 7. Let the arm run on its own
lerobot-rollout \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --policy.path=<your-username>/my-first-dataset --strategy.type=base

Verification status

Every command on this page was actually executed in the chenyang310/xlerobot:jetson image on JetPack 7.2:

Item	Result
Argument parsing for all 14 original `--robot.type` values	✅ Pass
Argument parsing for `--teleop.type` (including `keyboard`)	✅ Pass
`lerobot-calibrate` full chain	✅ Ran up to opening the serial port, then stopped (no servos were connected at the time)
`lerobot-teleoperate --teleop.type=keyboard` full chain	✅ Same as above
`lerobot-find-cameras opencv`	✅ Real hardware: detected `/dev/video0` and captured images successfully
`--robot.cameras` nested syntax parsing	✅ Pass
Original robots and XLeRobot coexisting in one process	✅ 8 types registered together successfully

No robot arm was connected, so motion-level verification hasn't been done yet. Once you connect your SO-101, walk through section 12 — tell me if any step misbehaves.

Next: troubleshooting/ — look here when something goes wrong

1. Environment Setup

2. Serial Port Check

3. Arm Calibration

4. Teleoperation

5. Camera Teleop

6. Data Collection

7. Model Training

8. Model Inference

Basics

Extras & Advanced

1. Environment Setup

3. Serial Port Check

4. Arm Calibration

5. Teleoperation

6. Camera Teleop

7. Data Collection

8. Model Training

9. Model Inference

I2C Communication

Serial Communication

Original LeRobot Operation ​

0. When to take this path ​

1. The complete workflow ​

2. Setup: find the serial port and cameras ​

3. First time with new servos: write IDs ​

4. Calibration ​

5. Teleoperation ​

Move the follower with the leader arm (leader-follower mode) ​

No leader arm? Use the keyboard ​

With cameras for visual feedback ​

6. Collect demonstration data ​

7. Replay to check ​

8. Training ​

9. Real-robot deployment: run a trained policy ​

10. Full table of command-line tools ​

11. Robot types the CLI supports (full list) ​

12. A complete example (from zero to running) ​

Verification status ​

Original LeRobot Operation

0. When to take this path

1. The complete workflow

2. Setup: find the serial port and cameras

3. First time with new servos: write IDs

4. Calibration

5. Teleoperation

Move the follower with the leader arm (leader-follower mode)

No leader arm? Use the keyboard

With cameras for visual feedback

6. Collect demonstration data

7. Replay to check

8. Training

9. Real-robot deployment: run a trained policy

10. Full table of command-line tools

11. Robot types the CLI supports (full list)

12. A complete example (from zero to running)

Verification status