Rackhouse Founder Spotlight: Dr. Ashutosh Saxena, Founder of TorqueAGI
Building the Foundation Model for Real-World Robotics
Most robots today only work in controlled settings. They're brittle, overly reliant on massive datasets, and built for repeatable tasks. Even with better sensors and machine learning, most systems break down in the real world. They get tripped up by shifting cargo in rail yards, cluttered warehouse aisles, bad weather on farms, or chaotic port traffic. Changes in lighting, terrain, or unexpected obstacles still cause basic tasks to fail.
This is one reason automation hasn’t scaled as fast as expected. Demos look polished, but take robots out of the lab and performance falls off a cliff. The friction points haven't changed. Teams encounter them constantly, they're just working with more advanced technology that struggles with adaptation just the same.
Dr. Ashutosh Saxena has seen this pattern play out over and over throughout his career. Throughout his work as a researcher and founder bringing AI into physical spaces, he kept facing the same problem: robots were great at learning tasks, but terrible at adapting when conditions shifted.
That disconnect between what was promised and what actually happened is why he built TorqueAGI: to give robots an intelligence layer that can reason through messy, shifting real-world conditions and stay reliable without endless retraining or massive datasets.
From Research to the Real World
Saxena has spent his career pushing AI beyond the lab. At Stanford, he worked with Andrew Ng on early breakthroughs in 3D perception, helping robots estimate depth from a single image and perform physical tasks like unloading a dishwasher. His academic work earned wide recognition, including more than 100 publications, over 23,000 citations, and awards from the NSF, MIT Technology Review, and the Robotics Science and Systems community.
But research alone was not enough. Saxena wanted to build systems that worked in the real world, not just in controlled environments. That drive led to a series of ventures: ZunaVision, Brain of Things, Katapult, and Caspar.AI. Each explored different dimensions of applied AI. Across those efforts, a consistent challenge emerged. Systems could be programmed to perform tasks, but they struggled to adapt when conditions shifted. Every new environment required more data, more rules, more intervention.
TorqueAGI came out of that friction. Saxena set out to move beyond point solutions and build a foundation model for robotics. A system that could learn from limited data, understand context, and perform across environments. Not just once, but every time.
The Intelligence Layer Robots Have Been Missing
TorqueAGI rethinks how robots interpret and respond to the world around them. Instead of treating perception, physics, and action as separate silos, it unifies them into a single world model that can reason under uncertainty in real time. This architecture captures relationships in space and time, how objects move, how force is applied, what’s likely to happen next, and gives robots the ability to reason through dynamic environments on the fly.
That means a robot can predict how a box might shift during transport or adjust its grip to avoid breaking a fragile item. It doesn’t need constant retraining or hardcoded rules. It learns from context and adapts in real time.
Deployment is simple. TorqueAGI runs as a software module. No hardware changes. No retrofits. Just new capabilities unlocked: hazard detection, adaptive pathing, precise object handling. The model runs locally, with no reliance on the cloud. It’s fast, secure, and purpose-built for high-friction environments like agriculture, mining, warehouses, and logistics.
Robots That Don’t Break When It Counts
TorqueAGI’s solution is not theoretical. It is already operating in high-pressure environments. Robots running these models are active in rail yards and distribution centers, handling real tasks like cargo inspection and dynamic navigation. Early adoption across logistics and industrial sites shows that the system works.
Zero-shot models have improved success rates from roughly 60% to 95%. Some deployments have reached nearly 100% accuracy. Customers also report faster rollouts, with deployment times reduced from years down to weeks. The system is highly data-efficient, learning with a thousand times less data than traditional models, and built for scalable unit economics.
What GenAI Means for the Physical World
Generative AI has raised the bar for automation. The expectation is no longer scripted workflows, it’s real-time adaptability. But while software systems have grown more generalizable, most robots haven’t kept up. They still depend on rigid control systems and large training datasets that don’t transfer well between environments.
That gap is becoming harder to ignore. As labor constraints tighten and operational complexity increases, static automation isn’t enough. What’s needed is software intelligence that can adapt in the field, learn from context, and recover when conditions change.
TorqueAGI’s foundation model is designed for that kind of pressure. It gives robots the flexibility to operate in messy, unstructured environments without constant retraining. And as hardware continues to commoditize, it’s the intelligence layer that will separate systems that scale from those that stall.
Rackhouse’s Bet: Reliability Over Hype in Real-World Robotics
Most robotics startups sell autonomy but break when reality hits. Edge cases, shifting conditions, or unpredictable environments. Saxena had spent years building AI for physical environments where reliability isn’t a feature, it’s the baseline. That mindset shaped everything: the product, the team, and the go-to-market. TorqueAGI launched in railways and logistics yards, where failure isn’t tolerated.
That bet paid off, and early adopters in those areas saw the impact. TorqueAGI made existing fleets adaptive without hardware overhauls. Operators reported faster rollout, stronger reliability, and consistent performance, even with limited training data.
What set TorqueAGI apart wasn’t just perception or planning. It was how the system responded to uncertainty. Most robotics stacks fail when real-world inputs don’t match training data. Torque’s software identifies when confidence is low, evaluates alternate paths, and adapts in real time. That feedback loop makes the system robust under pressure, even in unpredictable environments.
This is the filter Rackhouse uses. Not hype, not vision, but execution under real constraints. Systems that prove themselves in high-friction environments signal readiness for broader adoption. That’s how new categories are built. That’s where Rackhouse invests.
From Deployment to Infrastructure
TorqueAGI is building the intelligence layer that makes robots usable, consistently, at scale. The proof is in the performance: faster deployment times, stronger results, and dramatically less dependence on endless retraining cycles or hand-picked datasets.
Now it's about proving they can go wider. TorqueAGI is expanding beyond logistics and rail, tackling sectors like agriculture, construction, infrastructure inspection, and mining. These are places where complexity is the norm, not something that occasionally pops up. The goal is simple: prove that one model can juggle multiple tasks, handle different conditions, and run smoothly on any hardware platform, no matter the conditions.
That’s what it will take to reset the standard for embodied intelligence. Not better robots, but smarter systems. Ones that can learn, adapt, and stay online when variables constantly change.
Q&A with Dr. Ashutosh Saxena
You’ve spent over a decade building AI systems for physical environments. What kept breaking, and why did a new intelligence layer feel necessary?
For years, the failure mode was always the same: robots could follow a script, but they couldn’t handle the world as it is—messy, dynamic, and full of surprises. You’d see beautiful lab demos, then watch the system collapse the moment something shifted: a pallet was slightly rotated, the lighting changed, or mud splashed onto a sensor.
Every deployment turned into a cycle of brittle patches. Add a rule for case A, tune a model for case B, rewrite the pipeline for case C. It never scaled, because none of these systems understood the environment, they were just reacting to patterns they’d seen before.
That’s when it became clear that robotics didn’t need another perception model or motion planner. It needed an intelligence layer that could reason: how objects relate in space, how forces interact, how scenes evolve over time, and what to do when confidence drops.
TorqueAGI was built to solve exactly that. Instead of teaching a robot “how to do a task,” we give it a structured understanding of the world so it can adapt when the world changes. That’s the only way to make robots reliable outside controlled environments.
What does it actually take to move from “robot that completes a task” to “robot that adapts in the field”?
It requires a shift from programming a solution to modeling how the world works.
A task-completing robot knows a sequence: “pick this, place that.” It works perfectly until something deviates from the expected script.
A field-adaptive robot needs to:
understand spatial and temporal relationships
detect when the environment has changed
reason about alternate paths
make decisions under uncertainty
Adaptation isn’t a feature you bolt on at the end. It’s an architectural decision. You need a model that unifies perception, physics, and action in one structure, and runs fast enough on-device to react in milliseconds.
Only then can a robot behave like a reliable teammate, not a brittle automation.
What’s a common assumption about robotics that you think will age poorly?
The idea that autonomy scales simply by collecting more data.
We’ve had 20 years of teams trying to brute-force reliability with bigger datasets, and the result is still brittle systems that fall apart in the real world.
The future will belong to models that are data efficient—that can generalize across environments with orders of magnitude less data, because they understand structure and context rather than memorizing examples.
In ten years, people will look back at “collect millions of hours and hope it transfers” the same way we look at hand-crafted rules today.
In designing the system, what were the non-negotiables? What did the model have to be able to do from day one?
A. Interpretable and physics-grounded.
Robots in real environments can’t rely on black-box behavior. The model must expose how it understands forces, contacts, object relations, and uncertainty. This is critical for safety, debugging, and trust.
B. Reasoning over brute-force data.
Real deployments can’t depend on collecting millions of edge cases. The model must generalize from limited data by understanding the structure of how scenes evolve and what actions are feasible, so it can handle tasks from crushed boxes to trailer hitch alignment.
C. Unified intelligence across skills, delivered modularly.
Legacy robotics breaks because perception, planning, and control live in silos. TorqueAGI unifies these skills within one graph for a consistent world model, while exposing practical, modular agents that plug into existing stacks without redesigns.
D. Cross-embodiment from day one.
Robots vary in cameras, grippers, kinematics, and compute. The model must transfer across embodiments—yard trucks, warehouse arms, humanoids—adapting instantly when hardware changes, without retraining.
These principles shaped a reasoning-first foundation model: interpretable, physics-aware, unified across skills, and deployable on any robot platform. Not a point solution, but an intelligence layer built for the real world.
What’s the one result or deployment moment that gave you confidence this could scale?
There was a moment in a rail yard when our zero-shot model went live for the first time. The robot encountered conditions it had never seen in training (rain, shifting shadows, reflective surfaces from tank cars) and it still performed with near-perfect reliability.
That deployment improved success rates from around 60% to over 95% instantly, without a retraining cycle. Operators who had been skeptical said, “This is the first time the robot just worked.”
That was the turning point. When you see a AI system adapt on day one, in one of the harshest environments possible, you know you’re not looking at a point solution—you’re looking at AI infrastructure.
And once that happens in rail yards, you can see the path to farms, construction sites, warehouses, and everything in between.