From One Prototype to a Pilot Fleet: Building AMR and Humanoid Hardware Without Tooling
Content in this Article
In week three of a customer pilot, a humanoid test unit takes a bad fall and the sensor mast housing — LiDAR window, camera stack, IMU mount — cracks along a bonded seam. If that housing had come out of an injection tool, the team would now be negotiating an engineering change against hardened steel, six weeks and five figures deep. Instead it was an MJF nylon print: the revision added a rib, thickened a wall from 2 to 3 mm, and was back on the robot in 48 hours. The fall became a design input instead of a budget crisis.
Fleets of ten to two hundred, geometry that will not sit still
This is the defining production problem of the current robot wave. AMR pilots ship in batches of 10 to 50; humanoid pilots are smaller still. Every deployment teaches something — a sensor moves, a bumper grows, a cover gains a service door — and the enclosure geometry revs monthly. Injection tooling assumes the opposite: geometry frozen, quantity certain. Betting tooling money on a design that is still learning is how hardware startups convert runway into scrap steel.
Walk any robotics trade show this year and the pattern is visible on the machines themselves: SLS and MJF texture on sensor turrets, printed cable spines, lattice-walled covers. Teams in the humanoid race have been open about running structural and housing parts additively through development precisely so a crash on Tuesday can become a revision by Friday.
Atlas walked on printed legs
The deepest public version of this philosophy is Boston Dynamics' Atlas. To cut the humanoid's mass, its engineers rebuilt the lower limbs as 3D-printed structures with the hydraulic lines and actuator cylinders embedded inside the load-bearing geometry itself — replacing an assembly of machined parts, hoses, and fittings with structure that is also plumbing. On a walking robot most of the power budget goes into swinging and stopping the legs, so removing limb mass and inertia pays out on every single step.
The design move that matters: printing did not just make Atlas's legs lighter — it made the whole robot revisable. Those famous videos of the robot being shoved, tripped, and thrown onto pallets were also structural test campaigns, and a printed limb meant a failed part could come back redesigned without stranding any tooling. Crash, measure, revise, reprint: the loop is the technology.
Most teams will never print hydraulic legs, and do not need to. The same logic applies at far lower stakes: the LiDAR housing, the camera bezel, the charging-dock guide, the cover panel. Polymer instead of metal, bureau printing instead of an in-house lab — but the identical principle of freezing geometry late, after the fleet has voted.
Housings, covers, and the 10-to-200 zone
Sensor housings reward printing twice over. Around LiDAR and cameras you want matte, dark interior surfaces to kill stray reflections — MJF PA12 prints gray-black natively, no paint needed inside. You want datum bosses so a swapped sensor seats repeatably, printed light baffles between emitters and lenses, and snap fits that survive service — all fine in PA12 at 2–3 mm walls. Where EMI matters, a conductive coating or adhesive foil inside a printed shell passes many a pilot's compliance test.
Cosmetic covers are the other family. MJF and SLS carry structural covers comfortably to 100–150 units per revision. When customer-facing finish matters, vacuum casting takes over: a silicone mold pulled off one printed master yields about 15–25 parts, the mold takes roughly a week, and polyurethane resins span rubber-like to glass-filled, paintable to a consumer-grade surface. A few molds cover a 10-to-200-unit fleet without a gram of steel.
| Process | Sweet-spot quantity | Tooling cost | First parts in | Best for |
|---|---|---|---|---|
| FDM | 1–20 | None | 1–3 days | Brackets, internal mounts, jigs for the build line |
| MJF / SLS PA12 | 10–150 per revision | None | 3–7 days | Sensor housings, structural covers, snap-fit enclosures |
| Vacuum casting (PU) | 20–200, cosmetic | Silicone mold, ~15–25 parts each | 1–2 weeks | Customer-facing covers, rubber-like parts, painted finish |
| Injection, aluminum tool | 500–5,000 | Roughly US$3,000–15,000 | 3–6 weeks | A stabilized part with real volume forecast |
| Injection, steel tool | 5,000+ | Roughly US$15,000–50,000 | 6–10 weeks | The production fleet, after the design has stopped moving |
The arithmetic of not tooling yet
Run the crossover honestly. Say an MJF cover costs US$80 and its molded twin would cost US$6 plus an US$18,000 aluminum tool: the break-even sits near 250 units of a single frozen revision. A 60-robot pilot fleet that revises the cover twice a year never gets there — each revision resets the count to zero and, if you tooled, writes off the mold. The printed route looks expensive per part and is cheap per learning cycle; tooling is the reverse. Until the fleet stops teaching you things, per-learning-cycle is the price that matters.
| Keep printing when | Start tooling when |
|---|---|
| The part was revised within the last 90 days | A full pilot ran with no structural change to the part |
| Forecast per revision is under the ~250-unit crossover | Demand forecast clears the crossover for that specific part |
| Units are configured differently per customer site | One SKU serves every deployment |
| Crashes and field returns are still finding failure modes | Failure modes are closed out and warranty data is flat |
The robot programs winning this wave treat manufacturing process as a dial that follows fleet maturity, not a decision made once at kickoff. Print while the design is learning, cast when the surface starts selling, tool only the parts the fleet has stopped arguing about. When your next revision is ready, uploading the housing to our quote portal returns MJF, SLS, and vacuum-casting pricing against the same STEP file.