lerobot_robot_axol

LeRobot plugin for the Almond Axol dual-arm robot.

Installing this package makes the Axol devices available to the stock LeRobot CLI tools (lerobot-record, lerobot-teleoperate, lerobot-replay, ...) through LeRobot's plugin auto-discovery — no changes to LeRobot itself.

The device classes live in the almond-axol SDK; this package is a thin shim that registers them with LeRobot at CLI startup.

Requirements

• The robot machine (typically the ZED Box / NVIDIA Jetson wired to the Axol) with CAN set up — see the installation guide.
• Python 3.12+ (the hosted Axol installer bundles 3.13), Linux.
• LeRobot 0.6.1 (pinned by almond-axol[lerobot]).
• Almond Axol SDK 0.1.x (the plugin imports SDK module paths and intentionally stays below the next pre-1.0 API line).
• For the zed camera: the ZED SDK and pyzed (axol zed.install).

Install

pip install lerobot_robot_axol

This pulls in almond-axol[lerobot] automatically. To install from source instead:

pip install "lerobot_robot_axol @ git+https://github.com/almond-bot/axol#subdirectory=plugins/lerobot_robot_axol"

If you already installed the axol CLI via the one-line installer, add the plugin to that shared environment in place, then put its LeRobot entry points on your interactive PATH:

sudo /usr/local/bin/uv pip install \
    --python /opt/axol/uv/tools/almond-axol/bin/python \
    --no-deps \
    "lerobot_robot_axol==0.1.1"
export PATH="/opt/axol/uv/tools/almond-axol/bin:$PATH"

Use the in-place command above rather than rebuilding the Axol tool with uv tool install --with: a tool rebuild exactly reconciles dependencies and would prune separately managed packages such as the VIVE Ultimate runtime or a JetPack-compatible CUDA Torch build. The hosted environment already contains the compatible Axol SDK/LeRobot dependencies, so --no-deps also prevents a plugin install from changing that running environment. An Axol update reinstalls the tool environment, so re-apply the plugin install afterwards.

Usage

lerobot-teleoperate \
    --robot.type=axol \
    --teleop.type=axol_vr

lerobot-record \
    --robot.type=axol \
    --robot.cameras "{overhead: {type: zed, serial: 41234567, stereo: true}}" \
    --teleop.type=axol_vr \
    --dataset.repo_id=${HF_USER}/my-dataset \
    --dataset.num_episodes=5

Teleoperation uses a VR headset: open axol.almond.bot in the headset browser and connect to the robot machine — see the teleop guide.

For the highest-rate data collection (out-of-process ZED capture, Jetson NVENC dataset encoding, headset video streaming), prefer the SDK's own axol collect-data, which is built on the same LeRobot classes.