TorqueAGI founder Ashutosh Saxena says the company’s AI is designed to help industrial robots adapt to changes using physics-based reasoning and a few dozen examples, rather than tens of thousands. In an interview with Robotics & Automation News, he describes a system intended to work across different kinds of machines, while acknowledging that integration and site-level validation remain substantial work.
Watch Desk analysis
What happened
Saxena says TorqueAGI’s World-Reasoning Model represents objects, machines and workflows in a spatio-temporal graph, with geometry and physics built into its reasoning. He contrasts this with vision-language-action models that map visual input and instructions to actions using large amounts of demonstration data.
For a new task, Saxena says TorqueAGI typically needs a few dozen examples. He says its model has run on robotic arms, mobile platforms, humanoids, tractors and other equipment. The company’s account is that knowledge of physical processes can transfer between machines, while each machine’s controls, limits and movement still need to be described separately.
Saxena also says the model can use operational data customers already have and run as a dedicated private model on their own machines. He cites one warehouse deployment involving 700,000 orders and a heavy-equipment comparison in which a model trained on 6,000 events matched a system built from seven million events over nine years. These are examples provided by the company’s founder, not independently established performance results.
Why it matters
If the approach works beyond those examples, reducing the data and adaptation required for each new task could make industrial robots more practical to deploy across changing products, layouts and machines. Keeping a model on-site could also appeal to operators who want their equipment data and model behaviour under their own control.
But the pitch is not plug-and-play robotics. Saxena says work remains to connect manufacturer control interfaces, calibrate sensors and ensure the system respects each machine’s safety limits. He says TorqueAGI validates behaviour in simulation and supervised site runs before unattended operation. Those details matter: a model that reasons well still has to fit the machinery around it.
Our read
This is an interesting alternative to the idea that robot intelligence mainly comes from scaling up demonstrations. The valuable claim is specific: TorqueAGI says it can reuse physical and workflow knowledge while adapting to machine-specific controls. The evidence offered is promising, but it comes from the company’s founder, so buyers should ask for independently measurable results on the job they actually need done. “A few dozen examples” is a strong opening line; the full-shift record is the harder one.
What to watch
- Whether TorqueAGI publishes independently assessed results for its cited deployments.
- How much site-specific engineering remains when the model moves to a different machine.
- Whether customers can validate safety and performance over sustained real-world operation.
Discussion spark: For industrial robots, would you prioritise a model that needs fewer task demonstrations, or one with independently proven reliability across a full shift?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.