Newshot

Building a human-sized drone with AI in 6 days

Original text(via HN AI)

01 Overview

Zeus, an AI agent, helped design a heavy-lift drone, bought key components, and provided instructions for the build. In six days, the drone delivered its first payload.

We wanted to test whether an AI agent could question requirements and make engineering choices that held up under review. Could it prepare usable instructions for vendors, manufacturers, and other engineers? Could it help find failures before they happened in the physical world?

We found Zeus to be a strong contributor in workloads that were highly meticulous and time-consuming. It ran the hand calculations and simulations for the structural frame (3.1, 3.2). It found parts, provided supplier links, and tracked orders (04). It also wrote the 34-page wiring book the team built from (5.1). After the flight, the team generated 50 drone designs from the same brief (2.1).

Its limits were in engineering judgment and communication. Engineers still had to decide when a requirement or rule should change. Zeus could trace its calculations, but it did not always carry critical assumptions into the build instructions (06).

1.1 The setup

The team outlined some boundary conditions: 6 days, a limited materials budget, and an unlimited token budget. The drone was flown under Part 107 as a small unmanned aircraft system, weighing less than 55 lb including the payload.

Three full-time engineers worked on the build. The team had broad engineering experience but little experience building drones.

1.2 Project setup

The whole project sat in one central Zeus project, with shared CAD assemblies, a BOM, calculations, and other files. Within the project, engineers could start sessions for different tasks. Users quickly came to anthropomorphize the agents, dubbing them "Project Manager" or "Manufacturing Engineer."

Every project, session, and knowledge base was shared across the entire team, allowing new users to onboard quickly into existing context.

Upon use, it was immediately clear that existing models had highly uneven capabilities. They especially struggled with:

  • Engineering judgment. Understanding why a rule or requirement exists, and when a different approach is justified.
  • Communicating assumptions. Making sure the conditions behind a calculation reach the people who build the part.

The work fell into three kinds:

  • Work the team would have skipped. Zeus made a 34-page wiring book, with every wire length measured from the CAD, in 38 minutes (5.1). Nobody on the team would have drawn it by hand.
  • Work done together. The payload bay went through five reviewed revisions with an engineer in about two hours.
  • Broader exploration. The team generated 50 drone designs from one brief after the build, each with its CAD, parts list, and calculations.

02 Design

2.1 Engineers can now prototype by choosing between dozens of designs

Concepts are cheap for an agent. The team generated 50 drone designs using the same project requirements, steering each toward a different layout, including a hexacopter, a coaxial octocopter, and a frame of folded sheet metal. Designs can also follow abstract constraints, such as "make a quad that can fold into the back of a car." Most designs came back in one to two hours, each with its CAD, parts list, and calculations.

Hexacopter design
Hexacopter design

2.2 Engineers could track cascading interface changes in an assembly

A large part of the project centered on negotiating interfaces between subsystems.

The team considered replacing a machined aluminum landing bracket with a thicker printed bracket, which required longer clamp bolts. A fit check found the bolt tips and tether eye intersecting the payload plate in CAD. At an engineer's direction, Zeus lowered the payload mount and reran the bolted-joint calculations.

Three days in, I cannot keep this entire drone in my head… But Zeus can.Build lead

Three days in, I cannot keep this entire drone in my head… But Zeus can.

Section through the early printed leg bracket: the tether eye bolt and clamp fasteners intersect the payload plate.
Section through the early printed leg bracket: the tether eye bolt and clamp fasteners intersect the payload plate.
The same section after longer hangers lower the payload plate below the tether eye bolt and clamp fasteners, leaving a visible gap.
The same section after longer hangers lower the payload plate below the tether eye bolt and clamp fasteners, leaving a visible gap.

2.3 Zeus can make direct edits in connected CAD software, but requires highly detailed instruction

Zeus connected to the team's Onshape document to inspect parts and edit the assembly. Engineers could ask it to import components, change their positions, and update the hardware around them. Zeus saved the changes in Onshape for the team to review. Over the week, one engineer’s chat saved 12 versions of the assembly this way.

Drone CAD assembly, Nov 1, 7:46 p.m.. Starting assembly, rev E, with simple stand-in parts.
Drone CAD assembly, Nov 1, 7:46 p.m.. Starting assembly, rev E, with simple stand-in parts.
Drone CAD assembly, Nov 2, 12:18 a.m.. McMaster fasteners mated on all four arms.
Drone CAD assembly, Nov 2, 12:18 a.m.. McMaster fasteners mated on all four arms.
Drone CAD assembly, Nov 2, 8:40 a.m.. Avionics laid out on the top plate.
Drone CAD assembly, Nov 2, 8:40 a.m.. Avionics laid out on the top plate.
Drone CAD assembly, Nov 2, 11:13 p.m.. Maker models for switch, buzzer and PWM adapter.
Drone CAD assembly, Nov 2, 11:13 p.m.. Maker models for switch, buzzer and PWM adapter.
Drone CAD assembly, Nov 3, 8:49 p.m.. Work moves to the team's 3D-printed branch.
Drone CAD assembly, Nov 3, 8:49 p.m.. Work moves to the team's 3D-printed branch.
Drone CAD assembly, Nov 3, 9:07 p.m.. Top-plate corners cut, payload plate removed.
Drone CAD assembly, Nov 3, 9:07 p.m.. Top-plate corners cut, payload plate removed.
Drone CAD assembly, Nov 4, 12:24 a.m.. After upright sleeves, packs lie flat again.
Drone CAD assembly, Nov 4, 12:24 a.m.. After upright sleeves, packs lie flat again.
Drone CAD assembly, Nov 4, 12:36 a.m.. Pixhawk turned 90°, straps moved out.
Drone CAD assembly, Nov 4, 12:36 a.m.. Pixhawk turned 90°, straps moved out.
Drone CAD assembly, Nov 4, 9:11 a.m.. SEPS switch moves into a printed holster.
Drone CAD assembly, Nov 4, 9:11 a.m.. SEPS switch moves into a printed holster.
Drone CAD assembly, Nov 5, 12:59 a.m.. Holster rebuilt after later branch edits removed it.
Drone CAD assembly, Nov 5, 12:59 a.m.. Holster rebuilt after later branch edits removed it.
Drone CAD assembly, Nov 6, 3:18 p.m.. Printed mail slot added under the hub.
Drone CAD assembly, Nov 6, 3:18 p.m.. Printed mail slot added under the hub.
Drone CAD assembly, Nov 7, 12:37 a.m.. Markus's payload bay replaces the mail slot.
Drone CAD assembly, Nov 7, 12:37 a.m.. Markus's payload bay replaces the mail slot.

Nov 1, 7:46 p.m.Starting assembly, rev E, with simple stand-in parts

However, Zeus struggled with detailed design without concrete instructions, diagrams, and annotations. It could miss the broader intent behind engineering decisions or produce unnecessarily complicated designs.

For the payload-bay swap, Zeus imported the replacement design into Onshape and removed the previous bay from the assembly. It moved the new bay up 1.6 mm to match the revised mounting plate, updated the fasteners, and checked clearances. It then saved the updated assembly as a new Onshape version.

Also, can you switch out the payload for Engineer 2's design?

It was built on an older bottom plate. He measured a 0.190 in plate with its underside at z −29.8 mm. Engineer 3 has since made it 1/8 in, with the underside at −28.175 mm.

The bottom plate is now 1/8 in instead of 0.190 in, so the bay moves up 1.6 mm and its roof sits flush on the plate.

The bay shape itself is unchanged.

03 Analysis and verification

3.1 Agents do best with ongoing records of past decisions

Zeus kept a mass tracker through 5 revisions. Every line identified the source of its number: the maker's spec, the CAD, an assumption, or values calculated from formulas based on textbooks such as Roark's and Shigley's. It saved each calculation as a card with its inputs, their sources, its assumptions, and its result.

Can you show all of the calculations for current draw to hover?

3.2 Engineers with little simulation experience could identify vibration modes

When a hand calculation could not settle a question, Zeus ran a finite element analysis using its suite of tools. It compared coarse and fine meshes, then checked the result against a hand calculation. On the arms, that turned up a problem nobody had asked about. The team asked whether the arms and legs were strong enough. They were, but the arms' fundamental frequency (the frequency at which the arms would vibrate uncontrollably) fell within the band of frequencies generated by the props at hover.

It had a great library of techniques and was able to choose the right one for the job. It didn't make the mistakes that I often make of setting up boundary conditions wrong, subtly misunderstanding the problem, or being totally in the dark about how to analyze a certain situation.Build lead

It had a great library of techniques and was able to choose the right one for the job. It didn't make the mistakes that I often make of setting up boundary conditions wrong, subtly misunderstanding the problem, or being totally in the dark about how to analyze a certain situation.

Chart of the printed boom's vibration modes for hollow, 40% infill and solid prints, against the two bands where the props excite the boom at hover. No print case clears both bands.
Chart of the printed boom's vibration modes for hollow, 40% infill and solid prints, against the two bands where the props excite the boom at hover. No print case clears both bands.
I dock Zeus points because I found it often over-analyzed. We would change something very small, and it would then check every possible case, including ones that haven't substantively changed. Zeus also lacked the “it will be fine” instinct that many great engineers I know have.Build lead

I dock Zeus points because I found it often over-analyzed. We would change something very small, and it would then check every possible case, including ones that haven't substantively changed. Zeus also lacked the “it will be fine” instinct that many great engineers I know have.

04 Sourcing

4.1 Zeus helps teams buy and track the parts they bought

This capability was one of Zeus's most beloved features. Zeus helped source parts throughout the build. It told the team what to buy, provided supplier links, and tracked deliveries. Team members could check the chat to see what had been purchased and when it would arrive, without asking each other. A human engineer handled a few orders, including the waterjet plates, while working with Zeus.

When a part could not arrive in time, Zeus found an alternative. For example, the Mauch current sensor had to ship from Germany and could not arrive in time. Zeus moved current sensing onto the power board, which had a built-in sensor.

This is Zeus' best strength BY FAR. It can immediately find all necessary parts and not forget anything.Build lead

This is Zeus' best strength BY FAR. It can immediately find all necessary parts and not forget anything.

36 parts in 5 supplier carts. Listed total: $1,933.72.

Zeus also flagged a possible duplicate purchase: the team had already ordered the same Remote ID module.

where can i get it in san francisco

Before buying another one: Engineer 3 already ordered a Beacon gen.2 from Dronetag on Nov 2. Dronetag quoted 3 to 5 days, so it should land around today or tomorrow. Do you need a second one, or would a tracking number for that order do?

Remote ID moduleAlready ordered

Nov 5 to 7due

Beacon gen.2

05 Assembly and testing

5.1 Zeus does not have hands, but can make instructions for those that do

Zeus made the work instructions the team printed out and built from. Before the bench test, the engineers asked for a wiring diagram with the correct lengths, connectors, and clamps. In 38 minutes, Zeus made a booklet with every wire length measured from the CAD. The booklet included an overview of the whole system and a page for each step. It listed the parts, tools, and cut lengths, with numbered instructions and checks to pass before moving on.

But Zeus could not yet help with the physical assembly. That is only a matter of time.

The documentation quality of Zeus is incredible. Zeus made detailed diagrams, gave thorough explanations, and kept track of every important decision that was made at any time.Build lead

The documentation quality of Zeus is incredible. Zeus made detailed diagrams, gave thorough explanations, and kept track of every important decision that was made at any time.

5.2 Zeus’s context and constant availability made it a strong project manager

Zeus kept one decision log, one mass tracker, and one parts list for the whole team. That made Zeus the fastest way to get up to speed. Team members often asked Zeus to read every chat and give them a status update.

The same record kept the team aligned in person. Team members worked from the same printed sheets at the bench. When someone joined a new part of the build, they asked Zeus for an update on that work.

06 Limitations and deficiencies

Zeus's analysis was a strength of the build. It chose useful methods, checked calculations, and kept results traceable. The limits the team encountered concerned engineering judgment and how assumptions reached the people making parts.

6.1 Engineering judgment

Engineering judgment includes understanding why a rule exists and when departing from it is justified. A requirement can reflect a physical limit, a manufacturing constraint, or a choice that no longer applies.

Zeus tended to pursue the stated requirements rather than question whether a simpler approach could remove them. Engineers still had to make that call. The difficulty was knowing when to change a rule and how to justify the change.

6.2 Communicating assumptions

Zeus could trace a result through its inputs and decisions. But an assumption in the analysis also needs to reach the manufacturing instructions. For one landing-leg bracket, the analysis assumed a solid part without making the need for 100% infill clear to the person printing it. The bracket was printed at 10% infill and broke too easily.

This was a communication failure between analysis and manufacturing. The required print setting needed to accompany the part. The example does not show that Zeus could not check its calculations; it shows why build instructions need the same assumptions as the analysis.

6.3 Review and responsibility

To engineer is machine; to take responsibility is human. An agent can produce work faster than an engineer can review it. More autonomy can increase that volume and the effort needed to understand it. Reading and checking the output can take longer than doing a task yourself.

That review burden is context fatigue. It grows when decisions, changes, and assumptions are difficult to follow. In hardware, physical tests can be slow and expensive, and a wrong assumption can affect someone's safety.

The shared project helped by keeping calculations, inputs, and decisions available for review. Engineers set requirements, questioned the results, and decided what to build. Zeus contributed analysis and checked the work. Those records supported review, but engineers still had to decide which physical tests were necessary and whether to proceed.

07 The capability to come

This was one build with a small team. It showed useful contributions in analysis, sourcing, and documentation, alongside errors and substantial engineering review. It does not establish that an agent can carry out a hardware project on its own.

The team developed two practices to get the most from working with Zeus:

  • Record every decision. Keep the reasons behind a design next to the design: who decided, what it was checked against, and what changed. An agent can only build on context it can read. A team that works with an agent leaves this record as it goes.
  • Use more than one model. The smartest model is not always the right one. A fast, cheap model that people keep using teaches a team more than a slow one they avoid. Different models are good at different jobs.

The Zeus team is building the product around both practices. If you are building hardware and want to try Zeus on a real project, book a call.

08 Acknowledgments

A special thanks to Tessa Pugh and Emily Di Lauro for leading the primary build. Thomas Godden, Andrew Chen, and Alex Xiong for their help as well.

Jim Belosic and the team at SendCutSend for hosting the flight in Reno.

Abtin Khayatian, Bob Cao, Stanley Shi, and the Pantheon team for letting us use their space and 3D prints, and for sharing their wisdom throughout the build.

Thank you also to Chris Anderson, Aidan Cantu, Sanjay Dastoor, Ricardo de Arruda, Aeden Gasser, Otavio Good, James Ho, Donald Jewkes, Scott Langille, James Lin, Markus, Pavitar Saini, Aidan Smith, Steven Somiack, Matt Stokes, Clo Wang, Jackie Yi, and Alpin Yukseloglu. Friends of the team gave advice on radio, flight control, structures, and manufacturing.

Try Zeus on an engineering project.

This drone went from a kickoff call to its first flight in 6 days. Bring a project from your own program, and the Zeus team will show you how an agent fits into your team's work.

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by HN AI
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗