@axisrobotics deployed a single sim-trained policy handling 500-800 distinct automotive parts for Lotus/Geely. The scale matters because each part requires different grip dynamics and motion profiles—automation complexity scales non-linearly with part diversity.
Their approach: simulate the production line, train sorting behavior in sim, then deploy to physical hardware. The critical bottleneck isn't the initial deployment—it's how much manual recalibration is needed when new parts enter the mix. That's where the data engine economics show up.
If retraining cycles are short and transfer learning is robust, this model could compress capex timelines in high-mix manufacturing. If not, it's just expensive custom automation with a sim wrapper.
Watch the iteration speed on new SKUs. That's the real moat.
Their approach: simulate the production line, train sorting behavior in sim, then deploy to physical hardware. The critical bottleneck isn't the initial deployment—it's how much manual recalibration is needed when new parts enter the mix. That's where the data engine economics show up.
If retraining cycles are short and transfer learning is robust, this model could compress capex timelines in high-mix manufacturing. If not, it's just expensive custom automation with a sim wrapper.
Watch the iteration speed on new SKUs. That's the real moat.