Building AI for Reliable Execution: Lessons From Industrial Robotics

What it claims

Standard Bots' position is that industrial manipulation works when the learned model has a narrow job, the tasks are short-horizon, and edge cases are fixed with small amounts of targeted in-situ data rather than more data overall. Jentoft's concrete claim is that in-situ interventions "can fix an edge case with a few dozen examples." The company also argues that "no model today is truly hardware-agnostic." Because it builds the arm, end effector, control system and AI, it says it can co-optimize low-level control with higher-level policies. The article notes that Skild, which is pursuing cross-hardware generalization, would dispute this.

How it works

• Task scope: Tasks are deliberately short-horizon so they meet production cycle-time and reliability requirements. Any part of a workflow that doesn't need learned behavior is programmed conventionally.
• Training vs inference: Training runs in the cloud. Inference stays on-prem because most factories and warehouses lack reliable internet, and uptime decides whether customers accept the system.
• Sensing path: Wrist cameras and other sensors send raw pixels and signals over internal gigabit Ethernet to edge GPUs. The GPUs produce action chunks that stream to low-level control.
• Data sources: Simulation is used where possible, plus real demonstrations recorded with a handheld touch teaching device, plus failure signals and human corrections captured in production.

flowchart LR
  D[Deployed robot] --> F[Failures + human corrections]
  F --> Q{Air-gapped?}
  Q -- yes --> X[Data stays on site]
  Q -- no --> C[Cloud fleet training]
  C --> D

How to use it

StandardOS provides APIs and SDKs. Developers can use whichever parts of the stack they need, including their own models; NVIDIA Cosmos is the example given. The page does not show endpoints, SDK names, languages or pricing.

Limits

• Simulation gaps: Tasks involving liquids, suction or cutting flexible material are hard to reproduce in current simulators, so they depend on real-world data.
• Bring-your-own-model integration: Today this means writing your own integration code. Easier data collection, training and deployment are only planned.
• Air-gapped deployments: Many defense customers run air-gapped, and nothing flows back from those sites, so they get no fleet-learning benefit.
• Missing evidence: No benchmarks, success rates or cycle-time figures are given. The "few dozen examples" claim and the hardware-agnostic claim are company assertions in an interview, not measured results.
• Missing model details: Model details stop at "low billions of parameters" and a perception backbone trained on more than a billion images. No architecture, training data composition or weights are disclosed.