Sensors drift, elastomers remember, clocks disagree, and sim contact never met a real fingertip. Tensiq instruments the boundary between models and physical contact: standardized tactile observations, characterized sensors, and evaluations you can defend.
Illustrative stream showing the TT contract semantics, not a live sensor.
Different stages hit different failures. Same instrument underneath.
You have a gripper, a sensor eval kit, and nobody who wants to own the parser. You want touch working in week one, not month four.
→ the data layer you need: consistency at scaleYou are about to spend serious money collecting contact-rich data. Whether that corpus is usable in two years is decided at ingest, not later.
→ where corpora go wrong you need: an external instrumentYour pipeline works. Your problem is everything around it: unexplained gaps, noisy rigs, results that don't replicate, claims you can't audit.
→ the evaluation suiteMotor vibration, inertial loading, cable tugs, and thermal drift all show up in the contact stream. A fixed filter either cuts real contact transients or lets structure-borne noise through, because the spurious component changes with pose, payload, and temperature. Separating contact from everything else requires knowing the sensor's non-contact response per unit. That is characterization, not filtering.
Elastomers relax under held load, respond differently on the second press, and change with loading rate. Your model is learning a mapping whose input is silently history-dependent, spending capacity on the sensor's constitutive quirks instead of on manipulation. None of it transfers when a pad is replaced. We invert the material physics before the data reaches your model.
Gains drift and nothing in the stream announces it. Data collected across the drift is inconsistent with itself, so the dataset encodes a contradiction, and success-rate changes get attributed to the policy when they belong to the sensor. Fleet calibration state is a live variable. We treat it as one: drift models, health flags, and re-characterization triggers per unit.
The question is invisible until a paper submission, a safety review, or a dataset merge makes it catastrophic. If sensor identity, calibration state, firmware, and processing weren't captured at record time, they are gone. There is no retrofit. Every TT observation carries its full source context from the moment it is written.
Sometimes the task is solvable from vision and touch is redundant. Sometimes the channel, as delivered, is genuinely uninformative: unsynced, hysteretic, unit-inconsistent across the dataset, so the model correctly learns to discard it. You cannot tell these cases apart without a clean channel to compare against. We give you the clean channel and the ablation protocol.
Saturations, dropouts, and glitches arrive as ordinary-looking numbers. A few thousand bad samples in millions will destabilize training in ways that look like optimization problems. Explicit validity flags exist so your dataloader can mask what the sensor already knows is garbage.
The gap arrives at the end of the pipeline, after weeks of training, with no decomposition. Contact model? Sensor response? Randomization ranges? One worn pad? You cannot debug a gap you cannot decompose. We measure sim and real tactile streams in the same format and report where they correspond and where they diverge, by contact regime.
In sim the policy sees clean solver forces. On hardware it sees elastomer responses through a normalization someone wrote two years ago. Randomization is supposed to bridge the two, but nobody measured whether the sim distribution covers the real one. Characterized per-unit variation turns guessed ranges into measured ones.
Vision at 30 Hz, tactile at 200 Hz to 1 kHz, proprioception at 500 Hz, separate clocks, unmeasured driver latency. Slip onset happens in tens of milliseconds, so a 40 ms offset means the model sees the consequence before the cause. Policies still train on this and quietly plateau. Nearest-timestamp matching in the dataloader cannot fix what the recording didn't capture.
Custom parsers, vendor SDKs, per-unit calibration scripts, glue code. Then you swap hardware or add a second vendor and do it all again. Supported sensors emit one deterministic contract, so the pipeline work compounds instead of restarting.
Is the policy brittle or is rig 9's fingertip worn? Without a characterized apparatus you cannot tell a 6-point improvement from sensor drift, and teams burn weeks chasing regressions that were a dying pad. We establish the noise floor of your rigs and gate eval runs on sensor health, so the numbers you compare are comparable.
Safety bounds and compliance claims rest on readings from sensors nobody characterized, with unknown drift. When a partner, reviewer, or regulator asks for the basis of the claim, there isn't one. TT-instrumented episodes make force-domain assertions checkable, with calibrated uncertainty, and the eval report is signed and auditable.
"Same object, same gripper" is not same preload, same friction state, same pad wear. Initial contact conditions were never measured, so nobody can verify the task was actually the same. Instrumented initial-condition records make "same task" a measurable statement instead of a hopeful one.
The person changes teams and the baseline is gone. A versioned eval suite with provenance-complete records means the measurement outlives the person who ran it.
Success rate over fifty trials is not an evaluation. It is a coin-flip experiment with the diagnostics thrown away. Everything below runs from logged data, no robot in the loop.
Tactile predictions scored against ground truth in Newtons at 1, 5, 10, and 50 step horizons, against persistence and linear baselines. Error in units, not in someone's latent space, so the number means the same thing next year.
Average error hides everything that matters. We score contact make, load onset, slip onset, and break separately: detection F1 and timing error in milliseconds. A model that nails mean force and misses slip onset is not a manipulation model.
If the 90% interval doesn't contain the truth 90% of the time, your risk-aware constraint is a lie. Coverage and calibration error, reported per contact regime.
Non-negative normal forces, tangential forces inside the friction cone, no force without contact. Violation rates as a regression suite, so physics errors are caught before they become training data.
Matched contact interactions executed in simulation and on hardware, both emitting the same contract. The report says where sim tactile corresponds to real and where it diverges, by force range and contact regime. MuJoCo and Newton first, Isaac Sim next. Your sim gates become trustworthy because their trust region is measured, not assumed.
A short self-test against a reference fixture before every eval run. Pads outside their characterization envelope are flagged and excluded, so rig variance stops masquerading as policy variance.
Sensor identities, calibration states, initial conditions, suite version, all in the record. Reproducible by another site, citable in a paper, defensible in a review. Measurement conditions travel with the result.
The Tensiq Tensor is a deterministic tactile contract: physical units, explicit validity, uncertainty, and provenance, from any supported sensor, real or simulated. If you are adding touch to a robot, this is where you start. If you already run your own pipeline, the characterization and evaluation work above runs against your data either way.
Simulated contact emits the same contract as hardware, with provenance saying which is which. That is what makes sim-to-real correspondence measurable rather than assumed.
Data layer, metrology, evaluation. Each stands on its own: start wherever the pain is, keep your existing pipeline, and take only what earns its place. They compose well, but nothing here requires adopting the full stack.
Connect a supported sensor or bring your own; new adapters are scoped and built with your team. Sensor-specific complexity stays behind the interface. Your models consume one predictable contract.
Per-unit characterization of your fleet: gain, hysteresis, drift, non-contact response. Measured randomization ranges for sim. Health monitoring that flags a dying pad before it flags a phantom regression.
The eval suite above, run against your policies and world models, on your logged data and matched sim scenes. Signed reports your reviewers, partners, and future self can check.
You're training a VLA, diffusion policy, or multimodal robot foundation model and need a consistent tactile input layer.
You're building a world or action model and need contact observations aligned with vision, proprioception, and actions.
You're pretraining or post-training a multimodal model and need tactile data that aggregates across embodiments, sensor types, and collection systems.
You're building simulation or synthetic-data infrastructure and need real tactile observations to anchor contact behavior.
You're evaluating manipulation policies and need results that replicate across rigs, sites, and time.
You operate robot data-collection systems and need to know which tactile samples are valid, comparable, and reusable.
Tell us what you're building and where touch is fighting you. We'll map the fastest path to something useful: an integration, a characterization, or an eval pilot. A few technical details are enough to start.
Prefer email? info@tensiq.com. We'll use this information only to respond to your request. Please don't submit confidential datasets or proprietary files through this form.