What Eight Teams Built at Zeon’s AI Science World Model Hack
Brontë Kolar
On July 25, more than 300 people signed up to build with the latest AI, hardware, and bio tools at the AI Science World Model Hack, hosted at Zeon Systems. In 24 hours, eight teams turned ambitious ideas into working systems for real laboratory environments.
The point was not to stage a robot demo. It was to test how quickly a team could move from a scientific question to a workflow that could be previewed, reviewed, and run on physical hardware.
From a prompt to a physical experiment
Zeon’s platform begins with a world model of the lab. Depth cameras scan the benchtop, helping create a digital representation of the instruments, labware, and working space. That gives teams a place to preview their workflows in simulation before a robot moves.
To author a workflow, teams use a command-line tool that connects to coding agents. A natural-language prompt can become a robotic sequence, then simulation makes it possible to validate the sequence before deployment. Once it behaves as expected, the workflow can run directly on the physical system.
Three tracks, one shared challenge
Close the loop
The first track asked teams to run an end-to-end TEM-1 β-lactamase inhibitor screen: make the enzyme, verify it, screen compounds, and use the resulting data to decide what to test next. The challenge was as much about reliable experimental design as it was about automation.
Connect the lab
The second track explored the handoff between automated storage and the robotic bench. Teams worked with Openshelf and Zeon to make physical inventory accessible to a workflow—presenting an item, identifying it, picking it up, and using it without a person resetting the scene.
Build new physical skills
The final track tackled tasks that are difficult precisely because lab instruments were made for human hands. Teams worked on dexterity, perception, and verification—teaching robots not only to act, but to recognize whether an action actually succeeded and recover when it did not.
Eight teams shipped in 24 hours
Team MIKMAK. The overall winners built a human-supervised Design → Make → Test → Analyze loop for TEM-1 screening. Their system planned conditions, validated them in simulation, required approval before execution, and fed plate-reader results into the next round.
Hive Mind. This team connected automated storage, computer vision, and Zeon workflows to care for neural cell cultures. When a culture needed media, the system could retrieve it, replenish it on the bench, and return it to storage.
Tacit Teacher. The Track 3 winners used narrated videos and an agentic interview to turn tacit human know-how into a new robotic skill. Their ZPeel prototype tackled the difficult task of peeling sealed lab plates.
BugPicker Automation. By integrating an xArm with a specimen-handling and imaging setup, the team showed a path toward higher-throughput biological discovery with robotic picking, plate movement, and downstream analysis.
BioMate. BioMate built an explicit autonomous biology loop: design a plate and protocol, simulate with Zeon, execute on the robot, collect data, and analyze what to run next. The team carried two experiments through the system.
TeamFour. Using simple visual markers, the team made physical items legible to both the robotic bench and Openshelf, giving agents a way to find, retrieve, and store objects through a shared system.
Measuring Antibiotic Resistance with Color. This team built a closed-loop assay-development system that used agents for compound selection, workflow authoring, execution, and analysis. Their second round cut a baseline run from two hours and 45 minutes to one hour.
OT OneCapped. The team built a robust sample-prep demonstration with dual-arm cap removal, liquid handling, and visual verification at every step—using an original Opentrons OT-1 alongside modern camera-based checks.
What the weekend made clear
Simulation is no longer a consolation prize. It is becoming the development path: teams can author a workflow, check it in a living digital model of the lab, and then close the remaining gap to real hardware.
As coding and reasoning become easier to access, the bottleneck moves into the physical world. Teams repeatedly encountered the same questions: Can the robot find the right object? Did the gripper make contact? Did the cap actually come off? Verification is the work that makes autonomous execution trustworthy.
The strongest systems kept people in the design and decision loop. Scientists and automation engineers brought context that agents cannot infer from a prompt alone—steering the experimental question, reviewing simulations, and deciding what to test next.
This recap is adapted from Michael Raspuzzi’s original report, with a Zeon perspective on what the teams made possible. Read Michael’s full report