RoBoCo
Solves the problem of porting a legacy robotic operating program to a new cobot.
Core trick
Argon patches a ported robot program and keeps each patch only if its pass rate on a digital twin rises.
Description
Solves the problem of porting a legacy robotic operating program to a new cobot. Copying old code straight fails, due to unlogged failure cases and calibration gaps, and even a careful rewrite over weeks misses failure cases. Argon reads robot footage, telemetry and the legacy code in one 755K-token context, labelling failure cases, and then loops over a digital twin recreation of the robot. It watches each failure in a rendered video, patches the code, and keeps the patch only if the pass rate rises. In a simulated 1,000-cycle shift, bad boxes shipped drop from 256 to 0. Managed agents rebuild the twin in their own sandboxes and independently re-verify the results.
README excerpt
# RoBoCo (Argon-assistant Robot cell port) **Gemini 4 Argon ports a legacy ABB robot cell to new cobots, proves the port in a MuJoCo digital twin, and hardens it against the failures it found in real robot footage.** <p align="center"> <img src="progress/img/dash_4_loop.jpeg" alt="The Argon loop: what Argon saw, the patch it wrote, and the pass rate climbing" width="100%"> </p> <p align="center"><sub>The loop. Left: the twin clip Argon watched, with its failure pin. Right: the pass rate on the training and held-out suites, and the RAPID patch Argon wrote. Bottom: one card per iteration.</sub></p> | | | |---|---| | **10% → 100%** | Cell Pass Rate of Argon's port after 6 patches, chosen by the twin, not by hand | | **5 / 5** | held-out fault scenarios passed; Argon never saw them, and they never decided a patch | | **256 → 0** | bad boxes shipped as good per 1,000 simulated cycles, at the miss rate measured on real footage | | **755,591 tokens** | in one Argon call: 19 minutes of video, 218 telemetry logs and the legacy code, read together | | **4 / 4** | independent Gemini Managed Agents rebuilt the twin from public sources and reproduced every result | --- ## The problem A manufacturing automation engineer has to replace two ABB IRB 1200 robots, running 15-year-old RAPID code, with ABB GoFa cobots. The line can't stop for weeks of commissioning. The old program has never run anywhere except this cell. Nobody fully knows which assumptions it makes about frames, config