TorqueAGI's Saxena: World-Reasoning Model learns tasks from few dozen examples

TorqueAGI is positioning itself as a supplier of the intelligence layer rather than a builder of robot hardware, an approach that puts it alongside manufacturers instead of in competition with them. Saxena's argument is that most robot AI today is a language or vision model with robot data added to it, whereas TorqueAGI represents objects, machines and workflows through a spatio-temporal architecture with physics and geometry built into its reasoning.

The company's track record is still taking shape. It is an Nvidia Perception Partner and was selected for John Deere's 2026 Startup Collaborator Program, and it has worked with Dexterity on truck-unloading robotics. Saxena points to one large warehouse deployment where packing density improved across 700,000 orders, and to a heavy-equipment customer where a model trained on 6,000 events matched a system built over nine years and seven million events. Integration remains real work: Saxena said most remaining effort is on the machine side, connecting to control interfaces, calibrating sensors and respecting each machine's safety systems.

Saxena's case for keeping industrial AI private and local rests partly on regulation. The EU Machinery Regulation, applying from January 2027, treats AI-driven safety functions as high risk, and Saxena argues a machine cannot be certified on a model that cannot be versioned and tested. That sets up a trade-off he acknowledges: a shared model improves faster, but for critical equipment he argues ownership matters more.