CONTROLAgents & automation
JEV x Automation Room
Automata Room@automataroom
0:41This project explores giving an AI model, JEV, a programmable robotic environment to reason about tasks, allowing it to observe the consequences of different decisions in a simulated workspace.
SignalIf a component is outside the reachable workspace, the planner rejects it and explains why instead of pretending the robot can perform the action.
Design the cell. Watch the cycle time move.
We’re experimenting with JEV @typesafeai inside Automation Room to explore what happens when an AI model is given not just a prompt, but a programmable robotic environment.
The idea is simple:
Give JEV a robot. Give it a cell. Let it reason about the work.
A single sentence can generate a complete board layout components, connectors, memory chips, PMICs, and headers while the environment handles their placement without overlap.
From there, we can design the entire robotic cell:
Move the tray around the arm
Change the board position and distance
Adjust tool travel height
Choose between different robotic arms
Change the assembly order
Replan the operation instantly
Every change produces three candidate plans, while the 3D robot executes the same plan shown by the numbers.
The top down view makes the reasoning visible too:
reachability → tool paths → carried vs. empty movements → rejected positions.
And the assembly state updates live:
Waiting → Fetching → Carrying → Placed
The interesting part: the numbers matter.
The simulator shows that changing the assembly order doesn't dramatically change the total cycle time. The optimal ordering, calculated with Held–Karp, only saves around 0.5% in this example.
But changing the physical layout matters much more.
Moving the tray beside the board reduced the cycle time from 39.5s → 32.1s, roughly a 19% reduction.
Lowering the tool travel height can reduce it further.
So instead of simply telling an AI to “optimize the robot,” we can give it an environment where it can observe the consequences of different decisions.
Under the hood
The assembly simulator operates at centimeter-scale precision, separate from the larger room grid.
Motion is calculated per axis with acceleration profiles, while the arm trajectories follow the geometry of a fixed-base robotic arm that rotates and extends through space.
The same joint geometry is shared between the planner and the 3D rig.
That means a position accepted by the simulator corresponds to something the robot shown on screen can actually reach.
If a component is outside the reachable workspace, the planner rejects it and explains why instead of pretending the robot can perform the action.
This is the direction we’re exploring with JEV × Automation Room:
AI reasoning + spatial understanding + simulation + robotic execution.
Not just generating code.
Not just generating text.
Give the model a world it can reason about.