Bucket LabsA division of Bucket Robotics

SIM IS REAL

Trained on rendered scenes.
Detected real bricks.

THE REAL-WORLD TEST 98.2%brick detection F1 · 52 real photographs

The real world is just one sample. Research by Bucket Labs.

THE EVIDENCE / SIMULATION → REALITY

Trained in simulation.
Tested in the real world.

Four rendered scenes of loose toy bricks on varied surfaces
Synthetic training images
One real photograph of loose bricks with green detection boxes
Real evaluation photographs One available example with detection overlay; the aggregate score comes from 52 photographs.
0real photographs used for model training
98.2%brick detection F1
52real photographs evaluated

Detection at IoU ≥ 0.5 on scattered bricks. This score does not measure shape-and-colour identification or exact counting.

Evaluation note

Real photographs informed evaluation and development decisions; this was not a blind transfer test.

At the setting reported above, exact whole-scene counts were correct on approximately 40–46% of scattered photographs. Exact counting remains unreliable, particularly in piles. Correct totals can conceal cancelling misses and false detections.

The assembly films are visual demonstrations. Neither prototype recording has been tied to the evaluated checkpoint.

THE PROTOTYPES / WHAT COULD COME NEXT

From the model to the moment of work.

These films explore how digital parts and guidance could appear during a real build. They are visual prototypes, separate from the detection result above.

Digital assembly, part by part
Digital assembly overlaid on a hand Full recording
LEGO F1 / PROTOTYPE FILM

From instructions to the real build.

A 29-second walkthrough shows the digital assembly, guidance at the workbench, and the prototype’s on-screen step checks.

Edited prototype demonstration. Its on-screen checks are not a measured assembly-verification result; the model and run have not been linked to the brick-detection evaluation below.

From simulation to the point of work.

Research becomes useful when people can try it in their own environment. This iPad prototype proposes part identities and counts from a camera view.

The recording has not been tied to the evaluated checkpoint.
Watch the full recording