Helm.ai has disclosed $70 million in signed commercial contracts for its foundation models, according to The Robot Report. The figure matters less as a revenue number than as a signal: software trained to understand the physical world is moving out of research labs and into paid deployments. That shift has direct consequences for anyone evaluating a humanoid or industrial robot today, because the model running inside the machine is becoming as important as the machine itself.
What Helm.ai actually sells
Helm.ai builds what it calls foundation models for the physical world. Per The Robot Report, the company trains these models using an unsupervised methodology it calls "deep teaching," aimed at mastering the structure of the physical world itself rather than being hand-coded with rules for specific tasks. That general approach, learning structure instead of following a fixed script, is what separates a foundation model from the narrower, task-specific software that has run most industrial automation to date.
The $70 million contract figure, as reported by The Robot Report, is notable because it marks commercial adoption, not a pilot or a one-off demo. Capable demonstrations are easy to produce; getting paying customers to commit budget to a model they did not build themselves is a different and harder bar.
Why this is happening now
MIT Technology Review frames this moment as a new phase for industrial AI. For decades, industrial AI meant predictive analytics and other narrow, specialized applications bolted onto specific machines or processes, according to MIT Technology Review. What's changing, per the same analysis, is that advances in foundation models, physical AI, and agentic AI are making it possible to automate more complex industrial tasks than those earlier tools could handle.
Read together, the two reports describe the same shift from two angles: The Robot Report's reporting on Helm.ai shows one vendor converting physical-world AI into commercial contracts, and MIT Technology Review's framing shows the broader industrial sector naming that same category shift. Neither is proof the category is mature. Both suggest it is no longer purely experimental.
What this means for buyers
Most humanoid and industrial robot marketing still centers on hardware: degrees of freedom, payload, battery life, actuator torque. Those specs remain necessary. They are not sufficient. A robot's ability to handle a task it wasn't explicitly programmed for, adapt to a changed environment, or improve over time depends on the software layer, not the chassis. Two robots with near-identical hardware can behave very differently depending on which foundation model, or lack of one, sits behind the controls.
This creates a new evaluation problem for buyers. Hardware specs are measurable and comparable across vendors, which is why they dominate most spec sheets. Software and autonomy claims are harder to verify. "Foundation model," "physical AI," and "agentic AI" are being used broadly across the industry, and not every vendor backing them has Helm.ai's level of disclosed commercial traction or a published methodology. A vendor's willingness to specify what its model was trained on, how it has been validated, and which customers are paying for it in production is itself useful signal. Vague claims of "AI-powered autonomy" without that detail deserve the same skepticism as a keynote demo.
Practically, that means treating the "brain" question as part of due diligence, not an afterthought:
- Ask whether the autonomy or perception software is the vendor's own foundation model, a licensed third-party model (such as Helm.ai's), or a narrower rules-based system marketed with AI language.
- Ask for evidence of deployment, not just demos: paying customers, production environments, contract value where disclosed.
- Weigh that evidence alongside the physical specs tracked on Compare Robots' humanoid comparison and the pricing context on the humanoid robot price trackerrather than treating specs as the whole picture.
A category still forming
None of this means foundation-model claims should be taken at face value, or that Helm.ai's contracts settle which approach wins. The category is moving from demo to deployment, but it is still forming: methodologies differ, disclosure varies by vendor, and commercial traction at one company does not validate the field as a whole. Buyers should read it as a signal that the underlying software now warrants the same scrutiny long applied to actuators and batteries, not as a reason to stop asking questions. For ongoing coverage of specific platforms, see Compare Robots' robotics news and the best humanoid robots rankings, which weigh both hardware and software maturity.