AI should do the dangerous parts and the boring parts, and people keep the interesting parts. The biggest safety gains will not come from a humanoid doing something dramatic. They will come from thousands of quiet inspections that catch the failure before the shift starts.

How Should Engineers and Buyers Evaluate Robot AI
How Should Engineers and Buyers Evaluate Robot AI

Ashutosh Saxena, Founder | TorqueAGI

Tell us about yourself and your role with TorqueAGI.

I am Ashutosh Saxena, founder of TorqueAGI. I started working on AI for robots about 20 years ago, during my PhD at Stanford with Andrew Ng. Our first project was getting a robot to unload items from a dishwasher. The state of the art at the time was largely geometric, and what we found was that teaching the robot with data was a very powerful way to solve the problem. That insight shaped everything I worked on afterward. I went on to become a professor at Cornell, where we built RoboBrain, a project people called a Wikipedia for robots, where robots at different universities shared what they learned on a server and learned from each other.

As founder and CEO of TorqueAGI, my job is to build the AI that goes on enterprise robots and make sure it holds up in real conditions. What we build is a World-Reasoning Model, designed to understand physics, geometry and the way physical systems evolve over time, and the same underlying model runs across commercial tractors, mobile manipulators, construction robots, and humanoids.

 

Digital AI had fifteen years of internet data to learn from. What does the data problem actually look like for a robot on a warehouse floor or a piece of heavy equipment on a job site, and why can't it be solved by collecting more demonstrations?

Language models learned from the internet, which was fifteen years of text written by people for free. Physical AI has nothing like that. The data a robot actually needs is the near miss, the stopped line, the pallet that was wrapped wrong, the dust storm over a field. Those are exactly the events you cannot collect safely or affordably. You cannot ask a factory to stop its line ten thousand times so a model can learn what a stopped line looks like. You cannot crash a lift to learn how a lift crashes.

The industry's answer has been to collect harder: hundreds of thousands of hours of teleoperation, people paid to wear head cameras all day. I understand the instinct, but I think it is the wrong problem. The gap in physical AI is not a data gap. It is a physics gap. A new warehouse hire is shown a task a few times and is productive by lunch, because they already understand that objects fall, that boxes do not pass through each other, that a heavy load shifts when you turn. If a model understands that too, it can use the data a company already has, its operating procedures, its logs, its handful of demonstrations, and does not need a million more.

 

"Reasoning" and "world models" are everywhere in robotics right now. In plain terms, what does it mean for a robot's AI to reason about physics and geometry rather than memorise examples, and what changes for an integrator or OEM when it does, for instance in how many examples a new task needs?

Most robot AI today is a language model with robot data poured in. A language model is allowed to write a poem about flying to the moon and eating it. Nothing in it stops physically impossible plans, and a robot that reasons that way is dangerous on a production line.

A physical reasoning model works differently. It represents the world as objects, machines and workflows that exist in space and move forward in time, with physics and geometry built into how it reasons. So when something changes, it recalculates rather than searching its memory for a matching example. Tell it the gripper got smaller, and it works out the new grasp. Move the conveyor half a metre, and it adjusts. It can also explain why it chose an action, which matters when an engineer has to sign off on it.

For an integrator or OEM, the practical difference is in what a new task costs. With a model that memorises, every new SKU, cell layout or machine variant is another data collection exercise. With a model that reasons, a new task typically needs a few dozen examples rather than tens of thousands, and those examples can come from the customer's own operation. That changes the economics of deployment more than any single robot does.

 

A 90% success rate looks impressive on stage and is a recall on a production line. How should engineers and buyers evaluate robot AI, and what should they be asking vendors to report: success at line speed, over what period, in whose environment?

The public judges on video quality. The customer judges on whether the line stops. Those are different tests, and the industry has spent two years optimising for the first one.

A staged demonstration depends on an environment arranged to the millimetre. Even the more transparent companies publish laboratory success rates that would count as failure on a production line, because an industrial customer expects reliability in the region of several nines. Seventy-five percent in the lab is a proof of failure in the field.

If I were evaluating a vendor, I would ask for four things. First, success rate at line speed, not at whatever speed made the video look good. Second, over a sustained period, a full shift or a 24-hour window, because failure modes show up in hour six, not minute six. Third, in a customer environment, not the vendor's lab. Fourth, how the system behaves when it fails: does it stop safely, does it explain itself, can your engineer see why. I would also give credit to the companies that publish their numbers at all. At least you can hold them to something. The industry needs a shared definition of production-grade reliability, and buyers asking these questions will get us there faster than vendors will.

 

The same model can run on a picking arm, a mobile robot, a scissor lift and a tractor. How is that possible, and does it mean the debate over humanoids matters less for material handling and logistics than the industry assumes?

It is possible because the intelligence lives in the understanding of the workflow, not in the body. A model that knows how a warehouse flows, how a pallet is built, how equipment is allowed to move on a runway, can drive whatever body the job calls for. The body is one of the decisions about the operating environment, not the source of the intelligence. One model, many machines, rather than one robot, many applications.

On humanoids: yes, our model runs on humanoids, and there are industrial environments where a humanoid is the right body. But in most of the industries I care about, it is not. Humanoids cannot fly over a field to count crops. They cannot lift a pallet at 150 feet. For inspection, a legged robot is better; for heavy loads, something closer to a super-humanoid with a very different frame. The material handling industry already has decades of well-engineered machines. The question is not whether to replace them with something that walks. It is how to give the machines you already have the ability to reason.

So I would put it this way: the humanoid debate matters a great deal to companies that build humanoids. For a logistics operator, the more useful question is which form factor each task actually needs, and whether the intelligence can move between them.

 

Most robot AI today is one shared model that learns from every customer it is deployed with. For an OEM or a logistics operator, what are the trade-offs between that and a model that runs and learns privately on their own machines at the edge, in terms of operational control, accountability, and the testing and versioning that machinery regulation now requires?

When AI moves from generating information to controlling machines, ownership of the model stops being a data question and becomes an operational one. A digital model controls information. A physical model controls machines. The model determines how heavy equipment responds to an obstacle and whether a system considers a condition safe. If something goes wrong at a distribution centre or a gas plant, the operator has to be able to answer for the model that made the decision.

A single shared model that learns from every customer has one obvious advantage, which is that it improves faster. But it also means your operational knowledge, the way your facility runs, is feeding a brain that every other customer runs on, including your competitors, and that changes underneath you whenever the vendor updates it. That is the shared-brain problem.

We call the alternative Sovereign Physical AI: a common foundation, with each customer running its own dedicated, private model that lives on its own machines at the edge and learns locally under its own operating procedures. One foundation for many machines does not have to mean one shared model for every company.

Your machines, your data, your intelligence.

There is also a compliance reason this matters now. Industrial equipment has to be tested and certified end to end, and the EU's Machinery Regulation, which applies from January 2027, treats AI-driven safety functions as high risk. You cannot certify a machine on a model that is untested or that changes without you knowing. A private model that is versioned and tested per customer is not a preference in that world. It is a requirement.

 

The same perception that drives a robot can inspect a lift component or a vehicle underside before a person is at risk. Where do you see physical AI making the biggest difference to workforce safety, and how does that square with the labour shortages in warehousing and agriculture?

A meaningful part of what physical AI does today is not driving a robot at all. It is helping a person. Photograph a component on an aerial work platform and the model tells you whether it is damaged before someone stands on it at 250 feet. Check the underside of a vehicle on an assembly line without a person crawling under it. The same model that reasons about physics for a robot can reason about it for the worker standing next to the robot. Same intelligence, two customers.

On jobs, I think the honest answer is that these industries cannot fill the jobs they have. Construction  has millions of open positions in the United States. Warehousing has hundreds of thousands of vacancies and turnover above a third every year. An ageing workforce and a climate that is less predictable each season make heavy-equipment autonomy a resilience question, not just a productivity one.

So the way I see it, AI should do the dangerous parts and the boring parts, and people keep the interesting parts. The biggest safety gains will not come from a humanoid doing something dramatic. They will come from thousands of quiet inspections that catch the failure before the shift starts.

 

Dr. Ashutosh Saxena is the founder and CEO of TorqueAGI, a physical AI company building foundation models for enterprise robots in the industries that run the economy: logistics, manufacturing, agriculture and heavy equipment. TorqueAGI's models allow robots and machines to reason about geometry and physics, learn new tasks from a small number of examples, and run reliably at the edge inside each customer's own private model.

He has spent two decades at the frontier of AI for the physical world. His Stanford PhD research with Andrew Ng produced Make3D, one of the first methods for recovering 3D structure from a single image, an approach now standard in computer vision. As a professor of computer science at Cornell University, he authored one of the first applications of deep learning to robotics, recognised with the RSS Test of Time Award in 2023, and led RoboBrain, a large-scale knowledge engine that let robots learn from one another, described by MIT Technology Review as "the world's first knowledge engine for robots". His work has been cited more than 20,000 times.

He has co-founded four companies, including Katapult (NASDAQ: KPLT) and Caspar.AI, and has applied AI in healthcare and finance, where reliability and interpretability are prerequisites. His honours include MIT Technology Review's Innovators Under 35, a Sloan Research Fellowship, Microsoft Research and Google faculty fellowships, an NSF CAREER Award and the Smithsonian's "Eight Innovators to Watch". He holds a B.Tech from IIT Kanpur and an M.S. and Ph.D. in Computer Science from Stanford University.

 

 

The content & opinions in this article are the author’s and do not necessarily represent the views of RoboticsTomorrow

Featured Product

High Performance Servo Drives for localized and distributed control applications from Servo2Go.com

High Performance Servo Drives for localized and distributed control applications from Servo2Go.com

Engineered to drive brushless and brush servomotors in torque, velocity or position mode, Servo2Go.com offers a broad selection of servo drives in a wide range of input voltages and output power levels.