Home/Build one

B grade Authoritative secondary, traced to origin

The road to a humanoid: why it is not the first machine you buy

A humanoid is the destination of DIY robotics, not the starting line. This entry is about the road itself — which capabilities have to grow first, and what jumping straight to a humanoid really costs.

Route A

LeRobot Humanoid: about USD 2,500, joint motors + CAN FD

Route B

Berkeley Humanoid Lite: about USD 5,000, printed cycloidal gearboxes

Prerequisites

Desktop-arm chain working + one GPU that can run simulation training

Hidden costs

Spares and filament 10–20% of budget; sim compute; ~10 h/week upkeep in month one

MVP layers

Static stance → stepping in place → walking under one step/second

1. 1Full blueprints (LeRobot Humanoid)The best-documented open humanoid route; the $2,500+ bill of materials is checked in the entry
2. 2Joint motors (Robstride)Continuous torque and reduction ratio carry the decision; peak numbers set the order of magnitude only
3. 3CAN bus and driversTwenty-plus joints on one CAN bus; termination and wiring order are the two classic first-power-on traps
4. 4Printed parts (trunk and limbs)Load-bearing limb parts print along the load path; start with PETG
5. 5Host and computeStanding and stepping fit a Pi 5; vision policies are where the Jetson tier starts
6. 6Software stack: LeRobotCalibration, teleop and imitation learning in one place; the humanoid fork ships ready-made configs
7. 7Simulation (Isaac)A humanoid fall costs far more than an arm's; this is the robot class where sim pays back hardest
8. 8Policy models (OpenVLA)The gap between published success rates and real-hardware results is the field's least recorded fact

What counts as moving

Split the humanoid MVP into layers: ① static stance (balance inside the support polygon) ② stepping in place (centre of mass switching between feet) ③ slow walking (under one step per second). Projects like LeRobot Humanoid make ①② reachable out of the box (the official policy zoo gives you a start); ③ is where your tuning ability starts to matter. Set two expectations before anything else: standing is a start, not a finish line, and falling is not failure — humanoid tuning is grown from falls. It is precisely because a USD 2,500 fall hurts that the first two layers belong in simulation.

Parts and sourcing

The machines are covered in Blueprints: LeRobot Humanoid (USD 2,500; Robstride joint motors + CAN FD) or Berkeley Humanoid Lite (USD 5,000; printed cycloidal gearboxes + brushless motors, with labour as the main cost). Three hidden line items belong in the budget: spares and filament (falls consume 10–20% of budget), one GPU that can run simulation training (RTX 4090 class, or cloud by the hour), and your time — first-month upkeep on either machine is around ten hours a week, and that bill is not refundable.

Assembly order

The order follows the same principle as the quadruped at an order of magnitude more complexity: one leg on a fixture for homing and direction checks, then the second leg, then the torso. The real departure from everything before is power: CAN FD with brushless joint motors, not PWM signal wires — connected is not the same as communicating: termination resistors, common ground and bitrate must all line up at once. Print structural parts exactly to the drawing's infill and wall specs (80% infill, 5 walls is the floor for load-bearing parts); the hours saved there come back with interest the first time the machine tries to stand.

Software stack

This is the true dividing line between humanoids and everything earlier: the control policy has to work in simulation before it touches hardware. The chain: verify motions in MuJoCo/MJX or Isaac Lab → feed real parameters back into sim with the official identification scripts → deploy. LeRobot's policy zoo is a starting point, but the sim-to-real gap closes through system identification, not hand-tuning — builders arriving from servo robots tend to miss exactly this step, because hand-tuning used to work.

Bring-up to first motion

Order: stance in sim → power on hardware (suspended, strapped to a frame) → deploy policy → static stance → stepping in place → slow walking. Suspended power-up is not optional: a policy's first deployment thrashes, and the difference in repair cost between a thrashing robot on the floor and one on a frame is about tenfold. Three frequent traps: ① CAN packet loss — a joint suddenly goes unresponsive; check cabling and termination; ② skipping identification and hand-tuning instead — beautiful in sim, falls immediately on hardware; ③ incomplete logging — after a fall you cannot tell which joint failed first, and half of humanoid tuning is reading logs.

Where to go next

Three paths after MVP: collect teleoperation data and train manipulation policies (the LeRobot chain — manipulation is where a humanoid approaches practical value sooner than walking does); add a head camera and a VLA model (OpenVLA-class; consumer GPUs need quantisation); or go back into legged control research and train your own gait with RL — the point at which the 4090 finally earns its keep. Not recommended: designing your own humanoid structure from scratch — unless structural design is the goal, that is a different project, budget and timeline.

What we checked

We verified: the published BOMs, power architectures and printing requirements of both machines cross-check; that 'sim first, then identification, then hardware' is the process both projects document; and that suspended power-up is the consistent practice in community humanoid build logs. We have built no humanoid ourselves — the ten-hours-a-week figure and the spares percentage are community-scale, not our measurements. The grade stays accordingly.

What each route buys you

Sources

• LeRobot Humanoid and Berkeley Humanoid Lite official releases (BOM, power, printing requirements)
• Identification and training documentation of both projects
• Community humanoid build logs (suspended power-up, CAN faults, upkeep time)

Objects in this entry

• LeRobot HumanoidRobots & Kits
• Berkeley Humanoid LiteRobots & Kits
• Robstride joint motorsParts & Components
• CAN and CAN FD busesParts & Components
• MuJoCoSimulation & Toolchain

Last checked 2026-09-28