RobotyUpdate
The phone could track the room and still miss the table
Our early iPhone placement checks exposed different AR failure states. A visible camera-free target editor kept one useful action available.
Read postBuilding Roboty
Blog
Building Roboty and sharing what we learn along the way.
The local model loaded. The simulated arm never moved.
Two local Qwen3.5 35B workflow sessions stopped before simulated Reach execution. Their failures matter, while reaching performance remains untested.
Read postDeep dive
A cube's center can be in the tray while the cube is not
Our cube-to-tray evaluator checks the rotated object, support, release and settling. A worked edge case shows why a center-point test is too weak.
Read postBefore asking a model, we needed an arm that could reach a target
Our first simulated reach separated solving joint targets, moving actuators and scoring the tool point. Each could fail independently.
Read postUpdate
We started with an arm. Then we changed the first experiment.
The SO-101 gave us a concrete hardware target. A smaller simulation task gave us a way to separate setup, control and evaluation.
Read post