Prototype · runs in your browser
Same AI, same kit, twice. Once with Layer.
Both robots follow the same plans from the AI, pi0.5. The one on the right has Layer, deciding every 2 ms, live in this tab.
pi0.5 alone
0 of 4 fitted
- Gear
- Clip
- Bolt
- Bracket
pi0.5 + Layer
0 of 4 fitted
- Gear
- Clip
- Bolt
- Bracket
Loading the simulator and the cell (about 11 MB)…
It steadies the arm.
A new plan arrives about once a second. Alone, the arm jumps to it. Layer fades it in.
0 mrad alone 0 mrad with Layer biggest jump
It fixes the aim.
The AI's grab point is often a few millimetres off. Layer's wrist camera sees the part and shifts the grab.
–last shift
0 alone 0 with Layer parts fitted
It feels a bump.
Click either arm. Layer feels the push within milliseconds. Alone, nothing notices.
0pushes felt · 0 bumps from you
0.0 mm alone 0.0 mm with Layer knocked off its path
What it means for a cell
This kit. Alone, the AI gives up on the clip after 40.6 s. With Layer, all 4 in 26.8 s.
60 kits, same AI. Layer finished 12; the AI alone, none.
Same robot, same AI, same robot program.
About this model
What runs here
MuJoCo 3.12.0 in WebAssembly, 500 steps a second. Layer uses the chip’s integer arithmetic, within 0.4 mrad of our reference.
What is recorded
pi0.5’s plans and the wrist camera’s corrections, from a GPU run of this kit. Nothing replans after a bump.
Why this kit
One of the AI’s best: alone it fits 3 of 4 in only 5 of 60 kits. On average it fits 33% of parts alone, 64% with Layer.
Not yet proven
No board has run on a real robot. Here the simulation waits for each plan. Count the AI’s thinking time and Layer’s lead halves: 118 parts against 80.