Budget a Robot Learning Setup Beyond the Arm and GPU

Budget the next decision, not an imagined finished lab

Start by naming the question the proposed purchase will answer. A machine for testing data loading solves a different problem from a complete demonstration station. Write a deliverable for each budget stage: an environment that opens the selected data, a reviewed station that records usable episodes, or an evaluated baseline policy. This makes it possible to defer equipment that belongs to a later question without leaving essential current work unfunded.

The LeRobot installation guide, checked in its main version on September 5, 2026, separates core installation from optional workflow dependencies. Use that distinction when scoping a computing environment, but verify the exact release you intend to install. The editorial budget method here is deliberately price-independent. It asks teams to obtain dated quotes for a defined configuration and to document uncertainty, rather than treating a tutorial's equipment list as a complete procurement specification.

参考资料： LeRobot Installation, main documentation

Reconcile the complete hardware list

For every quotation, list what arrives, what must be assembled, and what is excluded. Depending on the chosen workflow, the station may require a teleoperation device, robot, power supplies, communication adapters, mounting hardware, cameras, task fixtures, and replacement consumables. Ask the supplier to identify substitutions explicitly. A similarly named kit may contain different printed parts, motor variants, or assembly services, so compare deliverables line by line rather than comparing headline totals.

The SO-101 guide links its bill of materials and distinguishes leader and follower motor arrangements. It is a useful reminder to verify the exact configuration; this guide does not prescribe interchangeable parts or undocumented modifications. Reserve qualified labor for reviewing assembly, power, mounting, tooling, and operating arrangements. Include delivery, taxes where applicable, inspection, and remedial work in the estimate. Record whether each line is quoted, assumed, already owned, or still unpriced before presenting the total.

参考资料： SO-101 setup and bill-of-materials reference, main documentation

Price observation as a system

A camera budget should cover the view you need to preserve, not simply a camera body. Include mounting, lighting where required, cable routing, connection availability, and time to check that the recorded scene answers the task's observation questions. Ask whether an existing camera can reveal contact and completion in the planned arrangement. Test that question in a reviewed pilot before buying extra viewpoints that increase data and integration work.

LeRobot documents camera backends including OpenCV and RealSense; backend support is not proof that every device configuration suits your task. Obtain the exact device identity and validate it with the intended operating system and software revision. Record the capture configuration that actually works. Budget a repeatability check after reconnecting the station, because a session that records the wrong view can waste more operator time than a modest mounting or identification improvement would have cost.

参考资料： LeRobot Cameras, main documentation

Separate training from runtime compute

List computing jobs independently: capture, inspection, training, evaluation, and artifact storage. A workstation that can display a camera stream has not thereby demonstrated that it can train the selected policy, and a rented training machine does not resolve local camera or robot connectivity. Identify which jobs can use existing resources and which require measured trials. Keep the policy configuration with any benchmark so the result remains meaningful when input sizes or batch settings change.

The installation documentation contains platform and accelerator-dependent setup paths. Confirm compatibility before committing to a purchase, and ask for a small representative run where possible. Measure whether the workload completes, its resource use, and the time needed to obtain a usable result. Do not convert a generic GPU specification into an unsupported training-time promise. For rented compute, include setup, data transfer, retained storage, and unsuccessful runs in the estimate, with an owner responsible for terminating unused resources.

参考资料： LeRobot Installation, main documentation

Estimate storage from a pilot and name every copy

Measure the output of a representative capture session before projecting storage. Include raw or retained source material, curated datasets, temporary processing output, model checkpoints, and backups. Specify retention periods and who may delete a superseded artifact. LeRobotDataset v3.0 combines encoded video, tabular data, and metadata, so a planning estimate should cover the complete dataset rather than video size alone. A missing metadata component can make an otherwise large archive unusable.

Worked hypothetical calculation: suppose a pilot produces 12 GB for one hour of recorded activity, and the plan calls for 20 recorded hours. The primary estimate is 240 GB. Retaining an original and an independent working copy brings that illustrative allocation to 480 GB before checkpoints, processing space, or backup overhead. These numbers are assumptions, not LeRobot performance facts. Recalculate after changing views, encoding, or retention policy, and distinguish recorded activity from the longer time spent running a session.

参考资料： LeRobotDataset v3.0, main documentation

Include labor and a reasoned recovery reserve

Estimate the work between successful demonstrations: preparing the scene, performing approved resets, inspecting episodes, explaining defects, and maintaining experiment records. Assign hours to those activities even when participants are volunteers. Otherwise, an apparently affordable equipment purchase can create a project nobody has time to operate. Track the first session's actual allocation so later estimates reflect your workflow instead of an optimistic recording-only schedule.

Build a recovery reserve from named failure scenarios. Consider a damaged cable, an unavailable operator, a storage failure, or a replacement part requiring renewed integration checks. For each scenario, record the likely interruption and the resource that shortens it. The SO-101 setup documentation is a technical starting point for component identity; it does not establish replacement lead times or your service capability. Obtain those answers separately. A reserve with explicit scenarios is easier to defend than a percentage added without explanation.

参考资料： SO-101 setup and bill-of-materials reference, main documentation

Worked hypothetical example: compare complete scenarios

Consider an invented budget expressed in planning units, with no currency or market-price implication. Scenario A allocates 40 units to hardware, 10 to observation and mounting, 15 to compute and storage, 25 to labor and integration, and 10 to recovery. Its total is 100. Scenario B advertises hardware at 30 units but needs 20 for observation, 20 for compute, 35 for labor, and 15 for recovery, totaling 120.

The cheaper hardware line therefore produces the larger project estimate under these assumptions. That does not prove either option is better: the assumptions need quotes and pilot evidence. Mark which entries are uncertain and identify the decision that could change them. A borrowed workstation, for example, only reduces the estimate after compatibility and availability are established. Compare equivalent deliverables, including who assembles the station and what evidence must exist before the next stage begins.

参考资料： LeRobot Installation, main documentation · SO-101 setup and bill-of-materials reference, main documentation

Attach evidence to purchase gates

Use a budget register containing item, purpose, owner, quote date, exclusions, confidence, and release condition. A release condition might be a successful loader check, accepted pilot recording, or approved integration design. This turns procurement into a sequence of reviewable commitments. When a gate fails, update the uncertainty and next action before moving money between lines. Avoid buying additional hardware merely to show progress while the central observation or task-definition problem remains unresolved.

Keep research and deployment budgets distinct in their purpose. A learning experiment can produce a valuable result without establishing production reliability, maintainability, or a validated safety design. Do not present those missing activities as minor extras after the research money is spent. Close the first budget cycle by comparing estimates with actual work and documenting reusable assets. The most useful saving may be a known-good environment or a reliable collection procedure that makes the next experiment easier to scope.

参考资料： SO-101 setup and bill-of-materials reference, main documentation · LeRobot Installation, main documentation

检查清单

• Tie every purchase to a defined experimental deliverable.
• Compare complete configurations and record quote exclusions.
• Include mounting, observation, integration, and approved reset arrangements.
• Validate the selected compute environment with a representative workload.
• Project storage from measured pilot output and explicit retention rules.
• Estimate operator, reviewer, and recovery time separately.
• Release later spending only after the preceding evidence gate is satisfied.

常见问题