Watch any robot demo from the last two years. Humanoids lifting totes, arms sorting parcels, quadrupeds climbing stairs. Now watch the hands. Every one of those machines is numb. It knows where its gripper is. It does not know what the gripper is touching, whether the part has started to slip, whether the surface is metal or foam, or whether the airflow across it just changed.
Humans run on this data. Your skin is not one sensor. It is four mechanical channels plus thermal ones, and most of them detect change rather than level. The receptors that catch slip and texture do not report pressure; they report how pressure is changing. You never sense that metal is cold. You sense heat leaving your finger, which is a rate, not a temperature. The feeling of wind is convective heat loss. Johansson and Flanagan's classic review of how the fingertips code touch for object manipulation makes the point in detail: skin is an instrument for measuring change, and evolution settled on that design for a reason.
Robotics calls this the missing epidermis layer. It is not missing for lack of sensor research.
The Research Exists. The Gap Is Where the Thinking Happens.
Tactile sensing is a mature research field. Optical tactile sensors such as GelSight and Meta's open hardware DIGIT image a soft gel to capture high resolution touch. Magnetic elastomer skins in the ReSkin and AnySkin lineage place a replaceable rubber skin over magnetometers and read three-axis force at hundreds of hertz, for tens of dollars. The BioTac fingertip proved a decade ago that pressure, vibration, and a heated core measuring heat flux are enough to discriminate materials the way a human finger does. Neuromorphic groups build event-driven skins that emit changes rather than frames, converging on the same design biology chose.
So the hardware question is largely answered, and answered cheaply. The open question is what reads the data. The field's dominant answer today is to ship the touch stream to a deep network: transformer representations over tactile data, trained on GPU clusters, deployed on GPU-class embedded boards. It is the same correlational architecture as the rest of AI, with the same costs. Heavy compute at the edge. Frozen models. No principled confidence, and no real answer to the question "how sure are you that this is slipping right now."
Why This Is Our Home Turf
Senua AI's core mathematics, causal states and epsilon machines, was built by physicists to find the intrinsic computation in a raw signal stream. Our whole architecture rests on one axiom: the atom of perception is the change, never the static value. We validated that the hard way on video authenticity. The same mathematics that produced only noise on static features separated real from manipulated footage cleanly once the atom became the frame-to-frame change.
Now look again at what skin is. It is a change-stream instrument. The sensor class biology converged on and the mathematics we run are the same shape. We do not need to invent touch cognition. We need to point the machine we already have at one more signal.
Concretely, and this is why the work ahead is mostly a software project:
- Slip detection, which is what friction perception actually is, uses our already validated anomaly gate on a new channel. We proved the discipline on UAV telemetry: learn the machine's normal regimes from its own stream, read a parameter-free confidence, and the gate closes when the process departs. Incipient slip is a departure in the vibration stream. Same gate, new wire.
- Texture and material identity are causal-state categories. The same mathematics that learned parts of speech and entity types from raw text learns which material a sliding fingertip is crossing, because the definition never changes: two histories belong to one state when they predict the same future.
- Thermal material identity is a heated element and a rate-of-change readout, the BioTac insight fed through the identical chain.
- Airflow is a hot-film element measuring convective loss. Also a rate. Also the same chain.
One machine, four senses, zero new cognition. The per-sensor engineering is a bus codec: a few hundred lines speaking I2C and SPI, written with the same discipline as the pure wire-protocol codecs we already run against drones, PLCs, and robot arms. No vendor SDK, no dependency, no model.
The compute budget is the punchline. A full tactile patch of thirty three-axis taxels at 400 Hz produces about 36,000 values per second. That is not a GPU workload. It is a rounding error for a CPU-native engine that already runs whole drone missions from an 8 MB binary. The commercial norm for robot perception is a GPU board costing over a thousand dollars. Our market research found no commercial deployment running full on-device autonomy in the no-GPU, small-memory class. Touch, on a computer that costs about sixty dollars, with the same engine that flies the drone and drives the PLC. That is the class we intend to occupy alone.
How We Will Prove It, and How We Could Be Wrong
Our development process requires the claim to be pre-registered before a line of engine code is written, with falsifiers that can kill it. For this arc they look like this. Slip events must close the gate on held-out strokes. At least eight materials must separate above chance through the standard chain. The heat-flux channel must discriminate material classes. A static-value control is expected to fail, and if it succeeds our change axiom takes a hit.
Ground truth comes from a self-labeling rig: a small motion stage drags the sensor skin across mounted material samples, so the rig's own motion commands label every stroke. No human annotation and no curated dataset. The same rig closes a loop we care about deeply. Senua commands the motion and feels the result. Perception and actuation through one substrate is the road to reflexes, such as grip that tightens the instant slip begins. That is the layer of the architecture, action from prediction, that we have designed but have always said, honestly, we had not yet built.
If a falsifier fails, we publish the failure and what it taught us. That is how our last three arcs went, and the negatives taught us as much as the passes.
Why Now
The market is turning physical. The newest startup cohorts are full of companies capturing physical-world signal to train models on later; I wrote about that turn, and its architectural blind spot, in The Physical-World Turn. We think the deeper opportunity is the machine that understands the signal at the machine, while it runs, on hardware that costs less than the sensors themselves. Skin is the next sense we give it.
One mind. Many bodies. Now, many senses.
References and Further Reading
The Biology: Skin as a Change Instrument
- Coding and Use of Tactile Signals from the Fingertips in Object Manipulation Tasks. Johansson & Flanagan, Nature Reviews Neuroscience 10 (2009). The canonical account of the four mechanoreceptor channels and their roles in grip and slip.
- Bayesian Exploration for Intelligent Identification of Textures. Fishel & Loeb, Frontiers in Neurorobotics (2012). The BioTac result: pressure, vibration, and heat flux discriminate materials at human-comparable levels.
- Robots with a Sense of Touch. Bartolozzi, Natale, Nori & Metta, Nature Materials 15 (2016). The case for event-driven, neuromorphic tactile sensing.
The Sensors: Open and Cheap
- GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force. Yuan, Dong & Adelson, Sensors 17 (2017).
- DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor. Lambeta et al., IEEE RA-L (2020).
- ReSkin: Versatile, Replaceable, Lasting Tactile Skins. Bhirangi et al., CoRL 2021.
- AnySkin: Plug-and-Play Skin Sensing for Robotic Touch. Bhirangi et al. (2024).
The Mathematics: Causal States on Sensor Streams
- Inferring Statistical Complexity. Crutchfield & Young, Physical Review Letters 63 (1989). The founding paper: causal states and ε-machines.
- Computational Mechanics: Pattern and Prediction, Structure and Simplicity. Shalizi & Crutchfield, Journal of Statistical Physics 104 (2001). Causal states as the unique minimal sufficient statistic.
- DisCo: Physics-Based Unsupervised Discovery of Coherent Structures. Rupe et al. (2019). Local causal states on spatiotemporal sensor fields.
- Computational Mechanics of Input-Output Processes. Barnett & Crutchfield (2015). The ε-transducer: the formal home of sensing and acting through one machine.
The Field's Current Reader
- Sparsh: Self-Supervised Touch Representations for Vision-Based Tactile Sensing. Higuera et al., CoRL 2024. The GPU-era state of the art this post argues one level beneath.
This piece continues the argument from "The Physical-World Turn": the industry captures physical signal to train models later; we build the layer that understands the signal at the source, while the machine runs. At Senua AI that layer already flies drones and drives industrial machines. Touch is next.