Standard Bots keeps its robot AI small and its motion logic in code

"We've built a zero-shot system that lets a user tell the robot which parts to look for on a given task," Leif Jentoft, Standard Bots' head of AI, told Latent Space, describing the company's machine-tending product ^[14]. The model handles perception, in his words "locating and identifying the parts," while "conventional programming handles the motion and the cell logic around it" ^[15].

Standard Bots describes itself as "America's largest AI-native industrial robot manufacturer" ^[1]. According to Latent Space, it recently raised $200 million at a $1 billion valuation in a series C led by General Catalyst and RoboStrategy ^[2]. Its customers include NASA, Amazon and Lockheed Martin ^[3]. Its arms are designed for machine tending, welding and assembly ^[4]. The company's homepage adds palletizing, pick and place, surface finishing and inspection, sold across four product lines named Spark, Core, Thor and Bolt ^[22][23]. The technical account that follows comes from Jentoft, co-founder and chief executive Evan Beard, and the Standard Bots website.

A small model, run on site

Beard said the company's largest model is in the low billions of parameters, which Latent Space noted is not large by frontier-lab standards ^[7]. Customers start from a shared base model and adapt it through demonstrations and fine-tuning, Beard said ^[5]. Jentoft said the company runs a range of models, including the zero-shot perception system used for machine tending ^[6].

Jentoft presented the size as a choice about data. "We believe data quality matters far more than raw volume," he said, "and we focus on getting the most out of targeted data rather than chasing the largest possible dataset" ^[8]. Beard tied the scope of the work to the factory floor. The company focuses on short-horizon tasks to meet production requirements for cycle time and reliability, he said ^[11].

Training happens in the cloud and inference runs locally, according to Beard ^[13]. "For factories, which is most of what we're doing right now," he said, "this is something where you need to do inference on-premise" ^[20]. Latent Space headed that part of the interview with sensors, edge GPUs and "action chunks" ^[21].

Those are two separate costs. Training is paid for in the cloud. Inference is paid for inside the customer's cell, each time the robot has to find a part, on whatever compute is installed there ^[13]. In our view the low-billions parameter count and the on-premise requirement belong together, though Beard stated them as separate facts ^[7][13].

Code owns the motion

Standard Bots uses conventional programming for the parts of a workflow that do not require learned behaviour, Beard said ^[12]. Machine tending shows where that line falls. Perception is the part that has to generalise. "Our backbone for this task is trained on over a billion images, which is what lets the robot distinguish between lighting conditions, material types, and an object versus its background," Jentoft said ^[16]. Motion and cell logic decide what the arm actually does, and they stay in code ^[15].

In other tasks the models go further. Jentoft also described models that generate actions feeding into the robot's control system ^[17], with "wrist cameras and other sensors" sending "raw pixels and signals into the system over internal gigabit Ethernet" ^[18]. Even then, the model's output lands in a control system Standard Bots builds itself. The company makes the arm, the end effector, the control system and the AI, Jentoft said ^[19].

Jentoft counts that ownership as the company's advantage. "That lets us co-optimize the models and the control policies," he said. "Despite claims elsewhere, no model today is truly hardware-agnostic, and co-optimizing low-level control and higher-level functions is a major advantage for both performance and iteration speed" ^[24]. Latent Space pointed out that Skild, which aims to build "general purpose robotic intelligence," might quibble, since its goal is generalisation across hardware ^[25].

Latent Space's earlier conversation with Applied Intuition reached a related conclusion. In that account, physical AI is constrained by more than model intelligence, and the hard part is deploying models onto real hardware under safety, latency, power, cost and reliability constraints ^[26]. At Standard Bots, that hardware is the company's own arm, end effector and control system ^[19].

The same split, outside the factory

We think this playbook carries over to software agents with little translation. A learned model is useful at the step where inputs vary: lighting, material, and where an object ends and its background begins ^[16]. The steps that change something in the world, where an arm moves or a cell sequence advances, are the ones Standard Bots keeps in conventional code ^[15]. For an agent working in software, the matching split puts the model on reading and classifying. The writes, the ordering and the retries stay in ordinary code that can be tested before it runs.

The company's other choices translate as well. It keeps tasks short to meet cycle-time and reliability requirements ^[11], runs a modest model where the work happens ^[7][13], and adapts a shared base with examples gathered in the deployment itself ^[5]. Jentoft treats customer access as the data advantage. "Our market access gives us unique access to the highest value data," he said ^[9].

A few dozen examples, measured how

"In-situ interventions can fix an edge case with a few dozen examples," Jentoft said ^[10]. It is the boldest data-efficiency claim in the interview, and the one a factory operator, or anyone copying the playbook for software, would most want to rely on.

In machine tending, the backbone underneath any such adjustment was trained on over a billion images ^[16]. Counting "a few dozen" generously as 100, the pretraining set is at least ten million times larger than the fix ^[28]. The two counts are not in the same unit, since an edge-case example may be a demonstration and not a single image ^[5]. Even so, the comparison separates two costs: a few dozen examples is the cost of adapting the model, and the billion images paid for a model that can be adapted that cheaply.

Read that way, Jentoft's two statements about data fit together. His preference for quality over volume applies to the adaptation step ^[8], and the backbone underneath it was built on volume ^[16].

In the interview, Jentoft did not say how a fix is measured: the failure rate before the intervention, the rate after, the number of cycles the result was checked over, or whether anything else got worse. The homepage, for its part, lists repeatability of plus or minus 0.1 Nm for the hardware ^[27]. A claim about a learned fix needs a number of that kind before an operator plans a cell around it. Beard has already named the yardsticks: cycle time and reliability ^[11].