← conru.com Collaborative Design with Autonomous AI Agents. Stanford, 1997. Read the thesis (PDF)
Cable harness routed by AI agents through an engine bay Open the simulator (works best on a laptop)
Live, right now A team of AI agents from my 1997 thesis is routing a cable harness in this window. Drag to orbit. Press ▶ Demo tour for a three-minute walk-through, or ⚙ Configure to change the problem.
⤢ Full screen Everything runs on your machine. Nothing is sent anywhere. What am I looking at? ↓
Stanford, February 1997

Thirty years ago I taught a team of AI agents to design things with a human.

This is my PhD dissertation, rebuilt so it runs in a browser tab. The window above is not a video. Creators, mutators and combinors are trading designs on a shared blackboard right now, and the harness you see is the best one they have found so far.

The thesis asked a simple question. Can a designer and a set of autonomous agents, working through a shared workspace, solve an engineering problem that neither could solve alone? The test case was routing cable harnesses through crowded 3D spaces, because it is hard, it matters, and it is easy to see when an answer is good.

The system was called CALIBER. It routed harnesses with up to sixty connectors and hundreds of wires, and designers at Lockheed tried it on real projects. Their early feedback: harness design time dropped from months to a few hours.

1997100 MHz SGI workstation. Today the same search runs in a browser tab, hundreds of times faster.
60connectors and hundreds of wires in the largest harnesses CALIBER routed.
~1066possible topological routings for the 20-port problem above (Table 2.2). Nobody searches that by hand.
Illustration: cable harnesses inside a satellite avionics bay
What the problem looks like in the real world: the avionics bay of a satellite, laced with harness. Every box needs wires to other boxes. Every wire shares a path with others for part of its route. Where the bundles split and merge decides weight, cost and whether it fits at all. (Illustration.)
The problem

Aerospace harnesses, and why they took weeks

A harness is a tree. Wires leave a connector, travel together in a bundle, split at a transition, and fan out to other connectors. A bundle of many wires costs far less per wire than wires run alone, so the whole game is deciding where to merge and where to split.

In the 1990s a re-route took weeks. The electrical requirements, the box geometry and the layout were all committed before anyone looked at the cables. Designers padded enclosures with extra clearance and hoped. My advisor Mark Cutkosky and I wanted the harness in the loop early, when the boxes could still move.

The plan after graduation was to make CALIBER a product for aerospace harness designers. The web arrived at the same time and I went that way instead. The ideas kept working, though. Search with a team of specialists, keep a human in the loop, share partial solutions on a blackboard. That is still how I build.

Figure 1.1 from the thesis: an eight-port harness routed through blocks and a cylinder
Figure 1.1 from the dissertation. An eight-port harness, its ports (P), transitions (T) and bundles, routed through three blocks and a cylinder. The environment in the live window is the same idea with twenty ports and a lot more clutter.
How it works

Four steps, all visible in the window above

Press Configure in the live view and open the Layers card to switch each of these on and off. The Demo tour walks through them in order.

Voxel wavefront growing from the obstacles
01 · CHAPTER 3

Map the free space

The enclosure becomes voxels. A wavefront grows out from every obstacle and wall, one shell at a time, recording how far each voxel is from anything solid.

Approximate medial axis graph of spheres and links
02 · CHAPTER 3

Find the corridors

Where wavefronts collide is the medial axis: the middle of every corridor. Maximal spheres along it, linked where they overlap, make a sparse graph the agents can search fast.

Agents' candidate routings drawn as a cloud over the graph
03 · CHAPTERS 4 & 5

Let the agents argue

Creators invent topologies. Mutators move branches. Combinors graft the good half of one design onto another. All of them trade through one design pool. The cloud is every idea in play.

Sphere and spring harness model in the designer's interface
04 · CHAPTER 6

Hand it to the human

The best design becomes a physical harness: spheres on springs, with tension, bending and obstacles pushing back. Grab a bundle and pull. Cut a branch and re-attach it. Then hand it back to the agents.

1997 screenshot: initial routing of a twelve-port harness
1997. CALIBER on the SGI: a twelve-port harness (Figure 6.13).
1997 screenshot: designer pulls a harness fragment
The designer cuts a branch and pulls it across the enclosure (Figure 6.15).
1997 screenshot: harness settles after the move
The harness settles into a new shape and the wires re-route (Figure 6.16). The same move works in the live window today.
Beyond cables

The same machine grows trees, rivers and lightning

A harness is a Steiner tree with economies of scale on its trunks. So is a telephone network, a transit map, a river basin, a lung. In 1997 the thesis closed with a note that the same agents should work in other routing domains. Now they do. Pick an environment in the live view, or click one below. Nothing changes but the scenery and the words. The agents, the cost model and the physics are the ones from the thesis.

Tree

🌳 Tree

Leaves, forks, branches. Leonardo's rule.
 Lightning

⚡ Lightning

One cloud, twenty strike points.
 Blood vessels

🫀 Blood vessels

A heart, organs, Murray's law.
 River basin

🏞 River basin

Springs, confluences, the sea.
 City transit

🚇 City transit

Stations, interchanges, lines.
 Data centre

🖥 Data centre

Racks, patch panels, fibre trunks.
 Plush creature

🧸 Plush creature

Paws, joints, limbs full of stuffing.
 Cable harness

🔌 Cable harness

The original. Engine bay, 20 ports, 200 wires.
What the experiments showed

Together beat either alone

A human working with the agents found better harnesses, faster, than the human alone or the agents alone (Section 7.4.3). No single agent was best in every situation. The team was.

Three ideas carried the result, and I still use all three.

  • Decompose. Break the problem into pieces that purpose-built agents can own.
  • Share. Every agent, including the human, reads from and writes to one design pool. No protocol. No hand-offs.
  • Evolve. Keep what works, recombine it, and let the human steer when the metrics cannot see what matters.

The dissertation is 204 pages: the routing problem and its search space, the environment representations, the blackboard, each agent, the designer's interface, and the experiments. It is readable, and the figures still hold up.

Read the dissertation (PDF)

Andrew B. Conru. Computational Support for the Collaborative Design, Routing and Manipulation of Cable Harnesses. PhD dissertation, Department of Mechanical Engineering, Stanford University, February 1997. Advisor: Mark Cutkosky. Supported by the NSF.

The question of how humans and autonomous agents work together is as live today as it was in 1997. It is the question I still spend my days on.
Andrew Conru · Open the simulator · Guided demo · Thesis PDF