A month ago I wrote SuperDex and the Missing Reader. Meta had released a contact-first physics engine for robot hands, and the part that feels had not shipped. I proposed tapping its contact stream with our causal-state engine and seeing what it made of it. We did. The question we set ourselves was simple: can Senua AI drive SuperDex, and can SuperDex simulate a robot hand holding a thin, slippery, fragile wine glass, so a real-time AI brain can be tested on the moment the glass starts to slip?
Everything is public: the rig, the scorer, the patch, the Senua harnesses, and every run we made, including the ones that did not work, in github.com/senuamedia/senua-tactile-reflex. Each hypothesis was written down before the run that tested it.
The Test
A two-finger hand pinches a wine glass with barely enough grip. The glass is carried and shaken, filled so the liquid's weight rises, and knocked with a 20 ms pulse. Some events stay within what the grip can hold; others exceed it. The hand has to tighten before the glass slides out, without squeezing hard enough to crack it.
Nothing physical is made up. The glass is the ISO 3591 wine-tasting glass, about 71 g empty. The fingertip is wet silicone on glass, friction 0.84. The crack limit, 2.89 N per finger, is derived from the strength of annealed glass through a thin-ring model. The human comparison is the measured grip reflex, 74 ms. Every value is tagged cited, derived or design in the parameter file.
Senua learns from 36 safe grips only. It is then tested on 16 unseen safe grips and 16 unseen over-capacity events, 8 pours and 8 knocks, with the same shapes as the safe ones, so nothing but the physics tells them apart. The fingertips report their forces and the glass's motion against the skin every millisecond.
The Glass Would Not Fall
On stock SuperDex the glass could not slip at all. In the over-capacity episodes it moved 0.012 mm on average, where the physics says it must fall (run_014). Before blaming anything, we needed to know why.
What Senua Saw
We pushed a block across a table in SuperDex, and applied the same pushes to exactly integrated textbook Coulomb friction. Senua read the sensor stream of each and induced its causal states: the smallest set of states that predicts what the stream does next.
| causal states | at rest | friction / normal force while sliding | |
|---|---|---|---|
| textbook Coulomb | 3: rest, and two slide states | 91% of moments | 0.600, the friction coefficient |
| stock SuperDex | 11: a ladder of steady creep speeds | none | rises with speed, 0.05 at 0.5 mm/s to 0.45 at 5.5 mm/s |
The ladder lies exactly on SuperDex's velocity-regularised friction law, the smoothing used by incremental potential contact: friction = μ(2x − x²), with x the sliding speed over 1 cm/s. Senua recovered the engine's hidden friction model from sensor data alone. In that model friction exists only while a contact moves, so a held object always creeps and a grasp can never hold still. Tuning the solver (iterations, tolerances, line search) left the creep unchanged; it was the friction law itself (run_004).
The Fix
Each contact now keeps a stick anchor, and friction acts on the displacement from it: the textbook elastoplastic treatment of Coulomb friction, with micrometre elastic compliance at the contact. A second defect showed up on the breakaway benchmark: the engine's approximate friction stiffness delayed breakaway, so with anchors on the exact one is used. Anchors are scoped to their scene. The change is off unless switched on, SuperDex is byte-for-byte unchanged when it is off, and it applies cleanly to the upstream commit (the patch and its notes).
| rotation-locked sled, 8 kg and 64 g | textbook | stock SuperDex | patched SuperDex |
|---|---|---|---|
| at rest while the push is below capacity | 97.4% | 17.4% | 97.3% |
| sliding friction / normal force | 0.600 | 0.600 | 0.600 |
| step at which a 1.3x-capacity push breaks away | 579 | 413 | 579 |
| breakaway on a 60 N/s ramp, x capacity | 1.118 | 0.672 | 1.116 |
| Senua causal states on the slip sensor | 3 | 11 | 3 |
The last row is the one I care about. Senua did not just find the fault; it gave the fix a pass/fail test. Physics that behaves like the textbook should have the textbook's causal structure, and after the patch it does (run_022). With it, the glass behaves as it must: safe grips hold, over-poured glasses fall, knocked glasses slip briefly and re-grip (run_023).
Senua at the Wheel
Then the part we set out to do. Senua runs as its own process. Every millisecond SuperDex sends the seven fingertip readings; Senua reads them as one chord, the joint change of all seven instruments, checks it against the causal machine it learned from safe grips, and replies hold or tighten. SuperDex applies the reply on its next physics step, so Senua alone decides when the fingers react.
Across 96,032 closed-loop exchanges, Senua's compute took 7.5 microseconds per moment and the pipe round trip 47 microseconds at the median, well inside the millisecond. The live decisions are identical, step for step, to a batch replay of the same episodes (run_025).
| reflex driven by | over-poured glasses dropped | knocked glasses dropped | safe grips dropped |
|---|---|---|---|
| no reflex | 8 / 8 | 0 / 8 | 0 / 16 |
| standard shear-rate reflex | 8 / 8 | 0 / 8 | 0 / 16 |
| human grip reflex (74 ms) | 0 / 8 | 0 / 8 | 0 / 16 |
| Senua AI | 0 / 8 | 0 / 8 | 0 / 16 |
Senua caught all 16 slips and, on pours, fired a median 4 ms before the glass had slid a millimetre. Its slide on pours matched the human reflex. It learned from safe grips only, with no threshold tuned by hand and no model of the glass given to it (run_024).
It did not get there first time, and the repository keeps every step. Reading one shear channel, Senua could not see a smooth slip at all, because a slipping glass's shear simply levels off like a steady hold. Reading the whole fingertip as an orchestra caught every one. A later readout did worse and was set aside. The simulated sensors are also perfect, so a simple rule on the one channel that measures skin motion directly gets close to the physical limit here. The test that separates approaches is real sensing with real noise, and that is where this goes next.
What It Adds Up To
- A physics defect found from sensor data alone, with the exact law responsible recovered, where solver tuning had changed nothing.
- A pass/fail test for the fix: the repaired engine must have the textbook's causal structure, and it does.
- A simulation driven in real time: microseconds per millisecond step, deterministic, learned from safe grips only.
- The glass kept in hand every time, pours and knocks alike.
The same engine that flies drones and watches network traffic read a physics engine's fingertips, found where the physics was wrong, and then held the glass. Next comes the harder half: Senua shaping the grip force itself rather than triggering a reflex, and real fingertip sensors with real noise.
References and Further Reading
The Experiment
- senua-tactile-reflex. Senua AI (2026). Rig, scorer, SuperDex patch, Senua harnesses, pre-registration and every run.
- Project SuperDex. Meta (facebookresearch), August 2026. Apache 2.0.
The Glass and the Hand
- ISO 3591:1977, Sensory analysis: Apparatus: Wine-tasting glass. International Organization for Standardization.
- Friction and wear characteristics of polydimethylsiloxane under water-based lubrication conditions. Lee, Sohn & Kim, Materials 15 (2022).
- Rubber adhesion and friction: role of surface energy and contamination films. Tiwari et al. (2020).
- Young & Budynas, Roark's Formulas for Stress and Strain, 7th ed. (2002).
- Signals in tactile afferents from the fingers eliciting adaptive motor responses during precision grip. Johansson & Westling, Experimental Brain Research 66 (1987).
The Physics and the Mathematics
- Incremental Potential Contact. Li et al., ACM Transactions on Graphics 39 (2020). The source of the smoothed friction law.
- Wriggers, Computational Contact Mechanics, 2nd ed. (2006). Return mapping for Coulomb friction.
- Compliance of elastic bodies in contact. Mindlin, Journal of Applied Mechanics 16 (1949).
- Inferring Statistical Complexity. Crutchfield & Young, Physical Review Letters 63 (1989).
- Computational Mechanics: Pattern and Prediction, Structure and Simplicity. Shalizi & Crutchfield, Journal of Statistical Physics 104 (2001).
This piece follows "SuperDex and the Missing Reader" and "The Missing Epidermis". The Senua AI engine is proprietary; its outputs, harnesses and every run are in the open repository.