Installing LeRobot: versions, environment and first checks

Choose the version envelope

Start an installation record with the operating system, processor architecture, Python version, LeRobot release and intended workflow. A workstation for offline dataset inspection needs a different dependency set from a machine controlling an SO-101. Name the first useful result, such as opening a local episode, before selecting extras. Keep experimental environments separate from a station already collecting valuable demonstrations.

The official release page identifies v0.6.1; its package metadata requires Python 3.12 or newer. Use Python 3.12 as the documented starting point here. The package permits a range of PyTorch releases, while the installation prose recommends a newer baseline for some paths. These are different statements. Record the actual resolved versions and do not combine commands from an older tutorial with this release without checking them.

参考资料： LeRobot release v0.6.1 · LeRobot v0.6.1 package metadata · Installation documentation at LeRobot v0.6.1

Create a clean environment

For an already installed and working Conda distribution, the tagged guide documents the following two commands, run separately: conda create -y -n lerobot python=3.12 and conda activate lerobot. Choose an unused environment name when an environment with that name already exists, and use it consistently. Installation of Conda itself depends on your platform and remains outside this recipe.

Confirm that the shell, editor and notebook kernel point to the same environment before diagnosing missing modules. Reopening a terminal can change the active interpreter. A notebook that still uses an earlier kernel can report an import failure even after the intended environment was repaired. Record interpreter location alongside versions; a package list without its interpreter identity leaves this common ambiguity unresolved.

参考资料： Installation documentation at LeRobot v0.6.1

Select dependencies for the first workflow

LeRobot v0.6.1 separates workflow dependencies into extras. For offline training and inspection, this guide adapts the documented package syntax with an explicit release pin: pip install "lerobot[training,dataset_viz]==0.6.1". Run it inside the selected environment after choosing a compatible PyTorch installation path. This pins LeRobot, but the resolver still chooses versions of dependencies within their allowed ranges.

Use the table to plan additions, then consult the tagged package file for their precise contents. A hardware SDK does not by itself provide the complete recording workflow. Conversely, a learning machine need not have motor access to train on recorded data. Record why an extra was installed so a later maintainer can distinguish a task requirement from an abandoned experiment.

参考资料： LeRobot v0.6.1 package metadata · Installation documentation at LeRobot v0.6.1

Resolve platform and video dependencies

Treat Linux, Windows, WSL and macOS as distinct environments, including processor architecture. A Linux recipe cannot establish native Windows support for a camera driver, and a working Windows camera does not establish access inside WSL. Begin with the exact workflow on the intended host. Avoid buying a new accelerator because a decoder failed before any model computation began.

The tagged installation page describes FFmpeg and TorchCodec setup, but its Linux ARM availability statement is inconsistent with the platform conditions in the same release's package metadata. Do not turn either passage into a blanket support guarantee. Identify the decoder actually selected, consult its matching dependency requirements and retain a successful local decode as evidence. FFmpeg being present on a shell path does not alone prove that the decoder can load the required libraries.

参考资料： Installation documentation at LeRobot v0.6.1 · LeRobot v0.6.1 package metadata

Verify the environment before hardware access

Run lerobot-info, the environment-reporting entry point included in v0.6.1, and retain the output privately with the installation record. Its implementation reports software and accelerator information useful for a reproducible issue. Review reports before sharing because file paths or machine details may identify your workstation. This check gathers environment evidence; it does not exercise a motor or certify device compatibility.

Next, open one authorised dataset episode in the matching visualizer and inspect a decoded image and a numeric feature. Finally, run a deliberately short offline workload appropriate to the planned policy. Keep the three results separate: imports succeeded, data decoded, and computation completed. A CPU check can help isolate data problems, but it cannot establish the timing or memory behaviour of the intended GPU workload.

参考资料： Environment information command at LeRobot v0.6.1 · Dataset visualizer implementation at LeRobot v0.6.1

Diagnose the first failing boundary

Classify a failure before changing packages. A resolver conflict belongs to dependency selection; a shared-library loading error belongs to the installed binary stack; a missing command may identify the wrong environment; a camera permission error belongs to device access. Preserve the first useful traceback and exact invocation. Repeatedly adding unrelated packages obscures the original condition and makes another machine harder to configure.

Worked hypothetical example: an environment imports LeRobot, but opening every video fails before the first training batch. The useful next comparison is the same local episode through the selected decoder, with its version and library error recorded. Buying more GPU memory or recalibrating an arm does not address that evidence. If decoding succeeds and the first batch instead exhausts device memory, record image sizes and batch settings before changing one workload parameter.

参考资料： Installation documentation at LeRobot v0.6.1 · Environment information command at LeRobot v0.6.1

Preserve a rebuildable installation

After the first checks pass, retain the environment export, package inventory, installation commands and any platform-specific choices. Record where a dependency came from as well as its version, especially when an accelerator wheel used a separate package index. An editable source installation also needs its commit and local changes; a release label alone does not describe modified source code.

Test reconstruction in a separate environment when the workstation becomes a shared project dependency. The acceptance question is whether another person can reproduce the declared offline checks using the record. Keep the working environment available during that exercise. When upgrading, create a new comparison record and repeat the affected checks before moving datasets or hardware sessions to it. Successful reconstruction is stronger evidence than a terminal screenshot ending with an installation success message.

参考资料： LeRobot release v0.6.1 · Installation documentation at LeRobot v0.6.1

Release the next stage deliberately

Finish with a capability statement precise enough for the next owner: for example, this environment decoded the named local dataset revision and completed the named offline workload on the recorded device. List unresolved jobs, such as camera capture or simulation rendering, individually. Do not let a broad label such as installed suggest those jobs were exercised.

Connecting a robot is a separate transition requiring the responsible competent person to review assembly, mounting, communication identity and operating arrangements. For a simulation-only project, keep physical device consumers disconnected and state that the result concerns software. Neither a successful import nor a completed training run establishes safe autonomous motion. The installation deliverable should make later calibration and training work easier to diagnose without overstating what has been demonstrated.

参考资料： SO-101 setup documentation at LeRobot v0.6.1 · Maintenance of work equipment

检查清单

• Record OS, architecture, interpreter location and LeRobot v0.6.1.
• Create an unused environment and select the same interpreter in every interface.
• Install the extras required by the declared workflow.
• Record the resolved decoder and accelerator dependency versions.
• Retain lerobot-info output and the first meaningful failure log.
• Verify an authorised episode and a short offline workload.
• Preserve rebuild instructions before upgrading or connecting hardware.

常见问题