Teaching robots to feel: How force feedback is reshaping physical AI | Talk w/ SenseGlove's Jordan Scholtes

Most robotics companies build one robot and hope it works everywhere. SenseGlove built a tool that lets humans teach any robot what it needs to know.

The distinction matters. While the industry debates whether robots need five fingers or six degrees of freedom, SenseGlove has been focusing on something more fundamental: how to transfer human skills to machines through force feedback. The company's latest product, the R1 (code name: Project Rembrandt), represents a point that the entire robotics industry is quietly making: from collecting vision data in simulation to gathering force data in the real world.

As Jordan, Head of APAC region at SenseGlove, explained during a recent conversation, the shift from VR training to robot teleoperation isn't just a product refresh. It's a fundamental change in how we think about teaching robots to manipulate objects.

SenseGlove is a Dutch company spun out of Delft University of Technology that makes force feedback exoskeleton gloves, hardware that physically resists your fingers so you can feel virtual or remote objects, not just touch empty air. The company spent years in the enterprise VR training market with its Nova and Nova 2 gloves, which were used by NASA, the Dutch military, and researchers.

But the R1 changes everything. It's built specifically for teleoperation and imitation learning. It collects real human force data to train robot hands.

Why force matters more than most robotics teams think

Vision tells a robot where something is. Force tells it how to interact with it.

This is the central insight behind SenseGlove's R1. The glove has 20 degrees of freedom and measures force across its fingers from 6.5 grams to 2.5 kilograms. That range isn't arbitrary. It's the difference between feeling the resistance of a sponge and the firmness of a bottle cap. These are two very different interactions that require precision, but would both look identical to a camera.

As Jordan demonstrated during the conversation, the glove pairs perfectly with robot hands equipped with pressure sensors. When the robot hand encounters resistance, that force is fed back to the operator through the glove. The operator feels exactly how much pressure the robot is applying, which means they can collect training data that's precise enough for autonomous operation later.

The most compelling example comes down to something deceptively simple: stocking shelves with cereal boxes. As Jordan explained, 

"If you squeeze it too much, what happens? It breaks. Or if you don't hold it well enough, it falls through your hand. So you need to hit that sweet spot."

That sweet spot, the exact amount of force required, is something a human instantly understands through tactile feedback, but it's nearly impossible to capture through vision alone. With the R1, an operator can demonstrate that sweet spot repeatedly, collecting "an immense amount of data" in the process. The robot then learns from that force data exactly how much pressure to apply every time.

The case for being in the middle

SenseGlove faces a compressed market. On one end, companies like Sanctuary use microfluidic force feedback for the most advanced haptics. On the other, open-source projects deliver basic tracking for under $600. SenseGlove sits deliberately in the middle by offering enough technology to reach your goal without unnecessary complexity.

According to Jordan, the philosophy is intentional: "Our vision is giving people enough technology to have sufficient power to reach their goal."

The R1 costs in the $20k+ range, which isn't cheap. But the company's argument is straightforward: can you get the tactile fidelity you need from cheaper hardware, or will simulation and camera-based learning leave gaps in your training data? They believe that without a glove capturing real force data right now, at this specific moment in time, many robotics teams are missing critical information about how objects actually feel.

As Jordan put it directly: 

"We do not think that getting all the most high level tech in our glove is going to necessarily help train robots or necessarily help make the VR training better."

Instead, the company focuses on having "enough information to reach your goals." The trade-off is deliberate. Texture sensors on individual fingers or exotic microfluidic systems might sound impressive, but they may not improve actual robot performance where it matters.

Whether that's true will likely depend on the task. A picking task in a warehouse may not need force feedback. A task that requires manipulation around delicate objects almost certainly does.

Where the language of interaction changes

Selling to VR customers is different than selling to robotics teams.

Jordan pointed out something that rarely gets discussed: the vocabulary changes completely. In VR, the "wow factor" can carry a project. A customer tries the glove, gets excited, funds an innovation project, and that's often enough.

In robotics, the first question is always technical. What's the exact measurement? Which robot hand are we using? What's the sampling rate? Does the data support autonomous operation, or are we stuck in teleoperation mode?

This is why the SenseGlove sales process now requires a robot engineer in the room, not just a business decision-maker. The glove isn't a consumer product or a training device anymore. It's a data collection instrument, and every specification matters.

As Jordan noted, "Everything needs to be predefined almost." You can't "riff" on a robot deployment the way you can with VR training. The data needs to be perfect because it's the code.

Cross-embodiment isn't a solved problem

One of the sharpest technical questions came up near the end of the conversation: how do you map human hand kinematics to robot hands with completely different structures?

My hand has different joints, different force ranges, and different pressure profiles than a two-finger gripper or a five-finger hand with servo motors. When I apply 500 grams of force with my index finger, how does that translate to something useful for a robot that grips completely differently?

Jordan's answer: 

“It depends on the pressure sensors. If the robot hand has accurate pressure sensors, you measure what the sensors are detecting and apply the same pressure distribution. The human operator becomes a proxy for the robot's own sensing.”

But there's no universal mapping. Every robot hand is different. Every pressure sensor operates at different speeds. Some measure at 100 Hz, others at 1 kHz. The integration has to be custom for every hand.

This is why SenseGlove is already integrated with Seed Robotics and has projects underway with Inspire, Robotera, and others. Not because the company is trying to be a universal standard, but because each integration teaches them more about what works and what doesn't.

As the robot hand companies develop hands with more degrees of freedom and better sensors, the SenseGlove becomes more useful.

Data collection at scale in Asia

One of the most revealing parts of the conversation was when Jordan mentioned what's actually happening in Asia right now.

Six-floor buildings with 500 to 1,000 employees just stock shelves repeatedly. All day. Collecting datasets. This is where the competitive advantage in robotics is being built, not in R&D labs or simulation environments, but in large-scale human teleoperation.

Europe and America face a different problem. The innovation is strong. The robot designs are cutting-edge. But the manpower to collect training datasets isn't there. Labor laws, aging populations, cost of living, they all make large-scale human data collection impractical.

This creates an interesting imbalance. Western robotics companies can design better robots, but they can't collect the data fast enough to train them. Asian companies can collect massive datasets, but they might be training less sophisticated hardware.

The companies that figure out how to bridge this gap, either through remote teleoperation, synthetic data that actually works, or novel training approaches, will likely dominate the next phase of physical AI.

What SenseGlove doesn't know yet

Toward the end, Jordan raised a question the company genuinely doesn't have an answer for: how much force is enough?

The R1 goes down to 6.5 grams. Do robot hands eventually need to sense even finer gradations? Do we need to go to 2 grams? Most robot hands on the market right now haven't caught up to what the R1 can measure, so the question is still open.

This is actually healthy. It means the hardware is ahead of the problem space, which gives teams room to experiment and discover new applications.

From teaching tools to data infrastructure

Over the next two years, SenseGlove plans to add a data business to its hardware business. The company might eventually license datasets the way major robotics companies do now, not just selling gloves, but selling the training data collected through those gloves.

For now, the focus is hardware first. Get the R1 integrated with as many robot hands as possible. Perfect the teleoperation experience. Make sure the force feedback is reliable enough that operators can trust it completely.

Then, once enough data exists and enough robots have learned from it, the company can move up the stack into the business of selling training datasets themselves.

Final thoughts

The metaverse hype cycle crashing was actually good for companies like SenseGlove. It forced customers to be concrete about what they needed. Instead of vague "immersive training" projects, companies now ask specific questions: "How do we train this robot to fold laundry? How do we collect force data from a grocery store restocking task? How do we move teleoperation to autonomous operation?"

Those are harder questions. They require better hardware and more precise integration. But they're also the questions that actually move the industry forward.

SenseGlove’s R1 exists because the industry stopped wanting the wow factor and started wanting the data. That shift, more than any individual product, is what will define physical AI in the next five years.

Related: Explore Senslove's R1 glove and force feedback hardware on Knoxlabs for the teleoperation and robot hand integration context.