Standard Bots is combining learned perception with conventional control software rather than asking one AI model to run an entire factory robot. Its arms handle jobs such as machine tending, welding and assembly, with customers adapting a shared base model through demonstrations and fine-tuning.
For a machine-tending task, a model identifies and locates parts while ordinary software manages motion and cell logic. The company says its perception backbone was trained on more than a billion images, helping it distinguish materials and objects under varied lighting. Its largest model is only in the low billions of parameters, reflecting a focus on targeted data rather than frontier-model scale.
Training happens in the cloud, but inference runs on GPUs at the factory. Cameras and other sensors feed the local system, which generates short action sequences for low-level controls. The company says this design avoids dependence on the unreliable connectivity found in many factories and warehouses.
Real deployments also provide failure signals and human corrections for later training when customers permit data sharing. Some defense installations are air-gapped, so nothing leaves the site. The architecture shows why industrial AI remains a hybrid engineering problem: learned systems can handle variable perception, while deterministic controls, local computing and customer-specific safety requirements remain essential.