Standard Bots bills itself as "America's largest AI‑native industrial robot manufacturer." In 2026 the company raised $200 million in a Series C round at a $1 billion valuation; the round was led by General Catalyst and RoboStrategy, a robotics‑focused fund. Customers include NASA, Amazon and Lockheed Martin.
We spoke with co‑founder and CEO Evan Beard and Head of AI Leif Jentoft to learn about Standard Bots' AI stack and how its models operate.
What the robot arms do
Standard Bots' industrial robot arms are designed for tasks such as machine tending, welding, and assembly. The company targets short‑horizon tasks that must meet strict cycle‑time and reliability requirements. Parts of the workflow that do not require learned behavior are implemented with conventional programming, while perception and identification rely on learned models.
Pretrained base models and task adaptation
Beard described a shared base model that customers adapt with demonstrations and fine‑tuning. Jentoft said the company runs a range of models, including a zero‑shot perception system for machine tending.
Their largest model is in the "low billions of parameters," by his account — not extremely large by frontier lab standards. Jentoft emphasized that Standard Bots prioritizes data quality over raw volume and focuses on getting the most from targeted data instead of chasing the largest possible dataset.
For the machine‑tending task, their backbone model was trained on over a billion images, which helps the system distinguish lighting conditions, material types, and separate objects from background.
Full stack and hardware–software co‑optimization
Standard Bots controls the full stack: the robotic arm, end effector, control system and AI. Jentoft argued that this lets them co‑optimize models and control policies, improving both performance and iteration speed. He also observed that no model today is truly hardware‑agnostic, so jointly tuning low‑level control and higher‑level functions is an advantage.
Other companies, such as Skild, aim for cross‑hardware generalization and may disagree with the hard claim about hardware agnosticism.
On‑prem inference: sensors, edge GPUs and "action chunks"
Training happens in the cloud, but inference runs locally on the customer site. Beard explained this is important because most factories and warehouses lack reliable internet, and uptime is critical for customer acceptance — especially for mobile robots.
Wrist cameras and other sensors feed raw pixels and signals into the system over internal gigabit Ethernet. Edge GPUs process that data to produce "action chunks," which stream to low‑level control.
Keeping the inference loop local also avoids operational issues associated with relying on cloud compute in production robotic systems.
Production failures, fleet learning and simulators
Standard Bots uses simulation where feasible, but Beard noted that some production tasks — such as those involving liquids, suction, or cutting flexible materials — are hard to reproduce in current simulators. Real‑world demonstrations and human corrections are therefore part of the learning loop.
The company captures failure signals and human corrections from deployments, but how that data is used depends on the customer. Many defense customers deploy in air‑gapped environments where nothing is sent back. For non‑air‑gapped customers, Jentoft said fleet learning typically delivers enough benefit to encourage data contribution in exchange for improved performance.
StandardOS: a robotics developer platform
Standard Bots has extended its platform to external developers through StandardOS, offering APIs and SDKs. Developers can build robotics applications using parts of Standard Bots' stack and can bring their own models. Today this requires developers to write their own integration code, but the company plans to simplify data collection, model training and deployment to its robots.
Beard pointed to NVIDIA Cosmos — an open family of omnimodal world models for physical AI — as an example, noting version 3 was released in late‑May.
Lessons from industrial deployments
Although not a flashy humanoid startup, Standard Bots is already deployed for industrial automation at organizations including NASA, Amazon and Lockheed Martin. The practical engineering lesson for AI in physical systems is to keep the model's role focused and use corrections from real deployments to address edge cases. The company's approach emphasizes targeted data and local inference as pragmatic responses to production constraints.
As Jentoft put it, and echoing recent remarks from Jev's creator Diogo Almeida, the key is "getting the most out of targeted data rather than chasing the largest possible dataset."
