The Robot Hand That Forgets How to Feel

Tactile sensing is the set of technologies that let a robot detect contact, pressure, and surface texture through touch rather than sight. Picture two robotic hands running the identical packing task on the same line, one wrapped in an electronic skin, the other built around a sub-surface ultrasound array. Both pass their acceptance tests on day one. Ninety days in, the first hand's contact readings have quietly drifted; the second hand's data looks almost the same as it did on launch morning.

That's the scenario UltraSense Systems laid out in a Robot Report op-ed this month, and it's a hardware story on the surface. Underneath, it's a data story. A sensor that drifts doesn't fail loudly — it keeps producing numbers, they're just wrong, and every training pipeline downstream inherits the error without a warning label. That's not a robotics-only problem. It's the same failure mode that shows up in head-mounted cameras, wearable rigs, and any field capture program that runs for months instead of an afternoon.

This post is about what a tactile-sensing durability argument teaches anyone buying, building, or collecting physical-world training data — egocentric video included.

What UltraSense Actually Said

On September 12, 2026, UltraSense co-founder Mo Maghsoudnia and CTO Hao-Yen Tang published a piece in The Robot Report making a specific claim: ultrasound sensing can provide tactile perception for robotic hands while avoiding wear and tear . Their target was electronic skin, the current default for robotic touch. The problem, in their telling, is placement. For high-duty-cycle robotic fingers and grippers, electronic skin faces a scaling challenge: the sensing layer is often located close to the harshest mechanical environment on the robot, an environment that includes repeated compression, abrasion, contamination, humidity, temperature variation, cleaning exposure, and material aging.

Here's the part that matters for anyone thinking about training data rather than hardware: over time, that can introduce wear, hysteresis, creep, delamination, baseline drift, and recalibration burden. None of those failure modes trip an alarm. They just make the numbers coming off the sensor slightly, then increasingly, wrong. For a laboratory prototype, these issues may be manageable. For a commercial robotic hand operating over millions of contact cycles, they become central to product viability. UltraSense's pitch is architectural: the better path is protected sub-surface ultrasound. Whether that wins the market is a hardware question. The data question is the one worth sitting with.

Wear Isn't a Hardware Problem. It's a Labeling Problem.

Every grasp a drifting sensor records still gets logged as ground truth. Nobody flags it, because nothing looks broken. Six months later, a policy trained on that data starts missing edge cases nobody can quite explain, and the postmortem takes weeks because the fault isn't in the model. It's three layers upstream, buried in a training set that was quietly rotting the whole time.

This is the exact bottleneck the broader robotics data market is racing to solve. As one recent industry account put it, unlike LLMs that were trained on a vast sea of publicly available text, robots need data that captures physical interaction, and that kind of data barely exists. That gap is creating a new kind of infrastructure business — companies whose entire product is making sure physical contact data stays trustworthy at scale. One such effort recently assembled a dataset described as containing 130,000 trajectories of robot manipulation data, 300 hours of simulation, and 100 hours of evaluations specifically to give researchers a clean, verified baseline. Scale alone doesn't solve drift. Verification does.

The Same Physics Shows Up in Egocentric Capture

Egocentric video refers to footage recorded from the wearer's own point of view, typically through a head- or chest-mounted camera, so the frame matches what a person or robot would see while performing a task. Consider a field technician wearing a capture rig for ten-hour shifts across a multi-week program: lens fogging, battery-swap gaps, strap slippage that shifts the frame a few degrees, IMU calibration that quietly drifts after the fortieth mounting cycle. None of that looks like failure on a monitor. It looks like slightly softer footage, a slightly off head angle, a slightly delayed timestamp. Multiply that across a hundred sessions and you have the exact same problem UltraSense describes for touch: contamination-by-use, not contamination-by-defect.

The scale at which this needs managing is already large. A recent egocentric dataset release reported 1,362 hours (80k episodes) of human demonstrations spanning 1,965 tasks, 240 scenes, and 2,087 unique demonstrators . That's thousands of individual capture sessions, each one a chance for the same slow-motion drift to creep in unnoticed. Different body placements carry different wear profiles too, which is a big part of why The Body-Zone Problem: Why Robot Data Collection Isn't One Job argues that a single capture spec can't cover a head-mounted rig, a chest rig, and a hand-mounted sensor with the same recalibration schedule.

What the Broader Data Market Is Telling Us

The volume side of this industry is moving fast. A recent industry report noted that the Open-X ecosystem has grown to over 1 million annotated robot demonstrations across 22 robot types , and separately that average data collection cost per hour fell 60% versus 2024, driven by teleoperation tooling and commodity hardware . Cheaper and bigger sounds like solved. It isn't. Volume tells you how much data exists. It says nothing about whether that data still means what it meant on day one of collection — which is precisely the drift argument UltraSense is making about touch sensors, just applied to an entire dataset instead of a single sensor. We've made a version of this case before, in One Million Datasets Later, Robotics Still Doesn't Have Enough Good Data: hitting a million demonstrations is a headline, not a quality guarantee.

Skilled human capture still isn't cheap, either. The same market report puts US-based teleoperator rates at $65–$120/hour for US-based teleoperators with domain expertise , competitive with domestic labor but well above overseas alternatives. Cost, volume, and drift resistance are three separate variables. Buyers who optimize for only one of them tend to discover the other two later, usually at the worst possible time.

Where This Leaves You: Four Ways to Get Field Data

So which option actually holds up once you account for drift, not just headline volume? There's no single right answer, but each path has an honest limitation worth naming.

• In-house teleoperation. Best for teams that already own the target robot and need data matched exactly to their embodiment. Where it stops: you're paying $65–$120 an hour per operator, per the figure above, and every hour is bounded by your own hardware fleet.
• Public egocentric datasets like Ego4D or the EgoVerse release. Best for cheap, fast pretraining priors before you spend a dollar on custom capture. Where it stops: none of it matches your specific tasks, environments, or camera placement, and none of it was collected under your quality controls.
• Licensed footage marketplaces. One such catalog is built from task demonstrations, tool manipulation, human-object interaction, open-world navigation, and first-person footage, sourced from more than 7,000 rights holders and cleared for AI training with documentation per asset . Best for volume with clean rights. Where it stops: it's still someone else's environment, not yours.
• Commissioned field capture programs. Best for matching a precise research question to a specific environment, task set, and quality-control schedule, which is the only real defense against silent drift. Where it stops: it's a program you build over weeks, not a download, and it costs more per hour than public data. We've broken down that math in Robot Training Data Finally Has a Price Tag.

None of these is universally correct. The honest answer is usually a blend — public data for breadth, commissioned capture for the environments and edge cases that actually determine whether a policy works when it leaves the lab.

The Real Lesson From a Touch Sensor

UltraSense's argument is narrow on its face: ultrasound beats electronic skin for durability. But the underlying claim generalizes further than they probably intended. Any sensor, camera, or capture rig that runs for months instead of hours will drift, and that drift shows up as bad data long before it shows up as a support ticket. If you're planning a training-data program — robotic, egocentric, or otherwise — the question isn't just "how much data do we need." It's "how do we know, six months in, that session four hundred still means what session one meant."

Frequently Asked Questions