Most robots in factories today work like actors who have memorized a single, precise dance. They can perform their steps flawlessly, but if you move the prop table an inch to the left, the whole scene falls apart. Devol Robots, a San Francisco startup founded in 2023, is betting that the next leap in industrial automation won't come from better choreography, but from giving robots a sense of touch.
Their approach, which they call "Physical Intelligence," trains AI models primarily on force, torque, and contact dynamics rather than just video [Perplexity Sonar Pro Brief, retrieved 2024]. The idea is to build a robot that can feel its way through a task, adjusting its grip in real time when a part slips or a bin is overfull, without needing a human to rewrite the code. It's a bet on moving from brittle scripts to goal-seeking autonomy, and it has attracted a $4 million seed round from investors including Y Combinator, Gradient, and Balaji Srinivasan [Announcing our $13m seed round from YC, pg, Gradient, Amjad, Guillermo, Balaji, and 120+ others | Mastra Blog, 2026].
The force-based wedge
Devol's technical wedge is its focus on force feedback as the primary sensory input. While most robotic manipulation research has been dominated by computer vision, the company's founders,researchers identified as Cheong and Ong,developed mathematical foundations for describing physical embodiment through force-based control [International Business Times, 2025 or 2026]. Their model integrates torque and force readings from each robot joint with visual data, using a recurrent neural network to predict physical interactions like acceleration and contact dynamics [International Business Times, 2025 or 2026].
The promise is a system that doesn't just see a wrench, but understands how it feels when turned against resistance. This physics-grounded model is designed to adapt in real time, allowing a robot to, for instance, gently wiggle a jammed component free or apply just enough pressure to seat a gasket without crushing it [International Business Times, 2025 or 2026]. For tasks in unstructured environments,think sorting a bin of mixed, delicate parts or assembling components with natural variation,this could be the difference between a useful tool and a very expensive paperweight.
The team behind the torque
Public details on the founding team are sparse, but the contours point to a research-driven origin. The company was co-founded by two individuals, Sze Yuan Cheong and Elijah Ong, who collaborated on the underlying research before commercializing it [International Business Times, 2025 or 2026]. Cheong, listed as CEO, is described as a serial entrepreneur with over a decade of experience in manufacturing and industrial engineering prior to moving into robotics [Most Robot AI Will Fail in Production, Here's Why - Grit Daily News, retrieved 2026]. This blend of deep technical research and hands-on industrial experience is a deliberate asset; knowing the math of force control is one thing, knowing the chaos of a real factory floor is another.
The company has grown to an estimated team of 18 people [International Business Times, 2025 or 2026], and is actively hiring. The early backing from Y Combinator and a roster of individual investors like Sarvesh Karkhanis suggests a belief in the technical premise, even at this very early stage.
An early map of the competition
Devol enters a crowded field of companies aiming to make robots more adaptive. The competitive landscape ranges from startups focused on specific applications, like Lucid Bots for pressure washing or Third Wave Automation for forklifts, to well-funded generalists like Figure and 1X pursuing humanoid forms. Devol's differentiation is its specific technical path,the force-based world model,rather than a particular robot form factor or niche task.
Seed Round (Mar 2026) | 4 | M USD
The company's stated goal is to enable "any robot to see, feel, and predict physical interaction" [Perplexity Sonar Pro Brief, retrieved 2024], positioning its AI as a potentially agnostic action layer. This is ambitious, suggesting future applications could span from traditional robotic arms to bimanual and humanoid systems. For now, the focus is squarely on industrial manipulation.
Where the wheels could come off
The ambition is clear, but the path from research prototype to reliable factory floor product is notoriously steep. The most immediate risk for Devol is one of traction and proof. The company reports it is "deploying its model with U.S.-based industrial clients" [International Business Times, 2025 or 2026], but no specific customer names, contract values, or performance metrics have been disclosed. In the cautious world of industrial procurement, a novel AI model must prove it is not just clever, but also more reliable and cost-effective than the simple, scripted robots it aims to replace.
- The integration burden. Deploying a new AI "brain" into existing robotic workcells, often from legacy manufacturers like Fanuc or ABB, involves significant systems integration work that Devol, as a small startup, may need to shoulder.
- The data moat. The force-based approach requires rich, physical interaction data for training. Gathering this at scale, across enough varied tasks and environments to build a robust model, is a formidable challenge.
- The economic equation. The ultimate question is unit economics: can Devol's system reduce the total cost of ownership for an automated task by enough to justify its complexity? The answer lies in reducing downtime, increasing flexibility, and cutting the engineering hours needed for reprogramming.
The company's answer, implied in its materials, is that the brittleness of current automation is itself a massive cost center. By building robots that can handle variation and recover from errors autonomously, they aim to unlock new classes of tasks that are currently too unpredictable to automate.
The next twelve months
The coming year will be about moving from early deployments to measurable scale. The $4 million in seed funding provides runway to refine the technology and pursue more pilot agreements. Key milestones to watch will be the announcement of a first named enterprise customer, the publication of any third-party validation of performance gains (like a percentage reduction in cycle time or error rates), and likely the beginning of a Series A fundraising process. The team will also need to demonstrate that its model can generalize beyond the specific tasks and robot models used in its initial pilots.
On the back of an envelope, the bet looks like this: if a single automotive assembly line spends 200 engineering hours per month tweaking scripts for a robotic pick-and-place station, and Devol's system can cut that to near zero while maintaining throughput, the savings could quickly run into six figures annually. That's the kind of math that gets a plant manager's attention.
Devol Robots isn't trying to beat the flashy humanoids at a game of dexterity. Its real competition is the incumbent logic of industrial automation itself,the belief that complex physical work must be broken down into millions of perfectly choreographed, fragile steps. Their success hinges on convincing manufacturers that a robot that can feel its way is smarter, and ultimately cheaper, than one that only knows its lines.
Sources
- [International Business Times, 2025 or 2026] The Founder Behind Devol Robots, the Startup Rethinking How Robots Learn | https://www.ibtimes.co.uk/devol-robots
- [Perplexity Sonar Pro Brief, retrieved 2024] Web-grounded brief on Devol Robots' product and positioning
- [Announcing our $13m seed round from YC, pg, Gradient, Amjad, Guillermo, Balaji, and 120+ others | Mastra Blog, 2026] Funding announcement referencing $4 million seed | https://www.mastra.com/blog
- [Most Robot AI Will Fail in Production, Here's Why - Grit Daily News, retrieved 2026] Article referencing founder Sze Yuan Cheong's background | https://www.gritdaily.com
- [LinkedIn, retrieved 2024] Sze Yuan Cheong - Co-Founder & CEO @ Devol Robots | https://www.linkedin.com/in/sze-yuan-cheong-b981a567/
- [LinkedIn, retrieved 2026] Elijah Ong - Co-founder & CTO | https://www.linkedin.com/in/elijah-ong-a2201a14a/