Standard Bots details AI stack for factory robots trained via demonstrations and real-world corrections

Standard Bots describes a shared base model that customers adapt through demonstrations and fine-tuning for factory tasks such as machine tending, welding, and assembly. The company’s largest model is in the low billions of parameters, emphasizing data quality over raw scale.

Its AI stack includes a zero-shot perception system for machine tending, with a backbone trained on over a billion images to handle variations in lighting, materials, and backgrounds. The model handles perception while conventional programming manages motion and cell logic.

Inference is performed locally on edge GPUs, with sensors like wrist cameras streaming raw pixels over internal gigabit Ethernet to generate ‘action chunks’ for low-level control. This on-premises approach is driven by factory requirements for uptime and unreliable external connectivity.

Failures and human corrections from real deployments are captured and used to improve models, where customers permit data sharing. Some defense customers operate in air-gapped environments and do not contribute data.

Standard Bots offers StandardOS, an AI robotics development platform with APIs and SDKs that allow external developers to build applications using parts of its stack, including bringing their own models. The company plans to simplify data collection, training, and deployment for developers.