Vision-language-action models have become a go-to method for training robotic manipulation, but a growing body of deployment evidence suggests they plateau well below the reliability thresholds necessary to be implemented across large projects. This can be seen in the field of precision manufacturing, where there can be no room for error and failure means destroyed product.
Sze Yuan Cheong, co-founder and CEO of Devol Robots, believes that to solve this problem means changing the underlying architecture, as vision-only pipelines never explicitly represent the physics of contact and control.
Devol Robots, the physical AI startup Cheong co-founded, is now deploying a force-based model to introduce a different learning program for robot learning.
The Problem With Predicting From Pixels
Consider a high-end optics manufacturer producing thousands of lens variants. Each lens requires a distinct jig and fixture, hand-tuned to accommodate its specific geometry. Tolerances swing from extremely tight to extremely loose depending on the variant. Even a well-trained human operator needs several minutes per lens and runs the risk of damaging it in the process.
The 6D pose of each lens at pickup is effectively unpredictable, as orientation varies with every cycle, and the placing position demands equal precision. With thousands of variants and hand-tuned jigs, nearly every cycle becomes an edge case.
This is exactly the kind of environment where standard vision-language-action models struggle. VLA pipelines learn to map what the robot sees to the next action, relying on visual correlations to infer the underlying physics of contact and control. Once they move from lab to industry environments, studies show these vision-first approaches effectively plateau well below the necessary reliability thresholds.
Competing solutions fall back on human teleoperators with VR headsets, manually taking care of cases the model isn’t equipped to deal with. But in precision optics, teleoperation is impractical. Operators can’t feel the tolerances through a headset, and repeated attempts compound the risk of lens damage.
“Their operators need to be really well trained,” says Sze Yuan Cheong, co-founder and CEO of Devol Robots. “The orientation is all over the place. You cannot predict the 6D pose of the object you’re trying to pick and place. There’s infinite variation.”
The Reason Behind These Failures
The failures in environments like precision optics trace back to a structural limitation in the way that VLA pipelines represent the world. In most of these architectures, robot behavior is learned by mapping visual observations and task prompts to action tokens with large, transformer-style models. The underlying assumption is that once they’re fed enough demonstrations and video data, the network will learn the rules of contact and control by itself.
The problem is that physics is never explicitly represented. Forces, torques, joint dynamics, stiffness, and compliance are not first-class state variables in the model. They’re inferred indirectly from pure visual context. The system learns correlations between what it sees and what action tends to follow, rather than modeling how the robot’s body physically interacts with the world. This means when it’s dealing with a task that requires precise force control across unpredictable orientations, that gap becomes the failure mode.
“In that pipeline, a lot of the context, like how we control the robot and how it interacts with the real world, alongside basic physics, is lost,” Cheong says. “It’s inferred by vision alone. We at Devol Robots don’t believe this is the right approach.”
Devol Robots: A Model That Reasons in Force Space
Devol Robots’ model proposes an alternative to the VLA pipelines it competes against. Alongside vision, it ingests torque and force readings at each joint plus stiffness parameters, learning how joints coordinate relative to the end-effector and how the full embodiment reacts to objects and surfaces during contact.
Training is structured around temporal sequencing using a recurrent architecture. Rather than mapping a static observation to a static action, the model analyzes the ways forces and motion evolve over time. Vision still plays a role in providing specific context, but it’s not asked to carry the burden of reconstructing contact physics from pixels alone. This means the model can plan trajectories based on physical feasibility, reasoning about why a trajectory should be planned a certain way instead of what trajectory was demonstrated.
In the optics deployment, that architecture translates directly into performance. The model can correct for jig variation, unpredictable pickup orientation, and tight placement tolerances in real-time. The robot effectively feels the object through force feedback and computes a feasible manipulation strategy as it goes along.
For the team at Devol, reliability should always be above 99%, compared with the roughly 50% that standard VLA models achieved in the same environment. “In manufacturing, you can’t fail,” Cheong explains. “You can’t say your success rate is 50% or 80%; you have to be better than 99%.”
The benefits aren’t confined to optics. In parallel e-commerce fulfillment trials, for example, the model picks and places diverse daily objects into bags and boxes without operator intervention, whereas those working with imitation-trained models typically need a teleoperator roughly every other minute.
Because control is abstracted into stiffness rather than exact joint angles, the model can generalize through different robot embodiments and tasks without per-embodiment post-training, a property that makes this process replicable in ways that vision-only pipelines aren’t.
Closing the Sim-to-Real Gap From the Bottom Up
Finally, one persistent bottleneck in robotic learning is the performance drop that comes the minute a model trained in simulation encounters the messiness of physical hardware. Because Devol Robots’ model is built on force experience, that gap narrows considerably. Force interactions in simulation closely mirror force interactions on a real robot in ways that pixel-based representations don’t.
“When we train in simulation, we don’t train the exact action,” Cheong says. “We train how each joint experiences force. This experience, this force-based control, is highly transferable to the real world.”
The team also uses imitation data differently from the mainstream approach. They extract the operator’s intent and let the physics-grounded base model execute it, with a pre-trained physics layer to make sure that if the model encounters a task it hasn’t seen before, it still produces stable motion.
The architecture amounts to a bottom-up inversion of the prevailing strategy in physical AI. Where most teams build top-down (trusting that enough vision data and compute will solve control implicitly), Devol Robots learns physics first and layers semantic reasoning on top.
Sze Yuan Cheong estimates robots will be capable of performing most human manipulation tasks within five years. But for now, through Devol Robots, he’s putting that thesis to the test straight into production floors, where failure isn’t an option and no human is standing by with a VR headset.
Disclaimer: GeekWire newsroom and editorial staff were not involved in the creation of this content..
