Physical-state inference for robotics
Robots do not just need faster AI. They need a trusted physical sense of their own body: which sensors are reliable, which regions are degrading, when contact is unsafe, and when learning should be suspended.
Tenbric HTDF is a local analog body-state layer. It turns continuous signals from force, torque, vibration, heat, current, strain and tactile sensors into compact, trusted state before the main AI stack acts.
The missing layer between sensors and AI
"The next robotics bottleneck is not only model intelligence. It is physical self-trust: knowing when the machine's own signals can be trusted before action or learning."
Too much raw sensing
Advanced robots push continuous force, torque, current, heat, vibration and tactile streams through scattered electronics and central compute.
No physical confidence
Adaptive systems need to know when a sensor has degraded, when a contact state is unsafe, and when unfamiliar data should block learning.
Local state inference
HTDF performs the first layer of physical interpretation locally, forwarding trusted body-state instead of continuous raw streams.
What HTDF gives the robot
Local reflex
Fast body-region response to contact, instability and unsafe physical change without waiting for central model inference.
Body-state inference
Continuous estimation of drift, fatigue, degradation, contact, heat, vibration, current and sensor confidence.
Integrity verification
Per-unit healthy baseline checking to detect when a sensor, actuator region or physical path no longer behaves normally.
Plasticity gating
Trust-weighted action permission, learning permission, escalation and derating signals for adaptive control stacks.
A physical compute layer for robotics
AI accelerators run models. Neuromorphic systems react to sparse events. MCUs and DSPs control deterministic loops. HTDF targets a different layer: continuous physical-state inference from noisy, drifting, multi-timescale sensor fields.
| Layer | Best at | Limitation | Tenbric position |
|---|---|---|---|
| GPU / NPU | Large-model inference | High power, centralised processing | Receives compact trusted state tokens |
| Neuromorphic | Sparse event processing | Less suited to continuous physical-state memory | Complementary sensor-boundary layer |
| Standard RC / ESN | Generic temporal feature projection | Uniform reservoir dynamics and weaker physical structure | Baseline Tenbric improves on |
| MCU / DSP | Deterministic control | Fragmented, software-heavy body sensing | Consolidates local interpretation |
| Tenbric HTDF | Physical-state inference | Needs silicon and OEM validation | Trust, drift, reflex, gating |
Structured substrate
HTDF uses multi-timescale cell bands, geometry-matched routing and coincidence/token features rather than a generic random reservoir.
Matched baselines
At matched cell count on audio benchmarks, HTDF beat vanilla ESN and LIF reservoirs by clear margins.
Chip-relevant
The substrate stays fixed; deployment adaptation lives in a small readout rather than retraining the reservoir itself.
Measured capability against explicit baselines
Simulation-side validation shows HTDF producing robotics-relevant capabilities: sensor fault localisation, graceful degradation, label-free unfamiliarity detection, and survival under fabrication-style non-idealities. The SHD result is a separate standard neuromorphic benchmark.
Sensor fault localisation.
Identified the degraded channel and fault type across 9 tested fault conditions: noise, stuck and scaled faults across 3 channels.
Cross-domain signal validation
The same baseline-and-drift procedure has been tested across public datasets spanning mechanical, thermal, electrochemical, chemical and compute systems.
| Domain | Dataset | Headline Result |
|---|---|---|
| Turbofan engines | NASA CMAPSS | 100/100 engines correctly ordered by degradation; 54-cycle median lead time. |
| Chemical process | Tennessee Eastman | 21/21 fault scenarios detected; 78 min mean detection time. |
| Thermal plant | LBNL boiler | 2.1x improvement over threshold alarms for fouling detection. |
| Electrochemical | CALCE lithium-ion cells | 4/4 cells tracked through degradation with actionable windows before fade. |
| Mechanical assets | Wind farm SCADA | 3/5 testable events detected, including +66h hydraulic fault lead. |
| Compute telemetry | Oxford Reveal V100 | 97.9% healthy classification; AUROC 0.742 vs 0.491 threshold baseline. |
Initial robotics wedges
Surgical robotics
Sensor trust, tissue-contact state, local degradation awareness and safety gating where physical uncertainty is commercially expensive.
Industrial cobots
Certification support for adaptive behaviour: contact response, corrupted-sensor detection and operating-state confidence.
Humanoid body regions
Hands, wrists, feet, joints and actuator clusters producing local body-state tokens for the central AI/control stack.
Prosthetics & exoskeletons
Adaptive control with user-specific drift, fatigue and fit-state inference.
Defence & aerospace
State trust under partial failure, spoofing, drift, sensor degradation and hostile operating conditions.
Industrial monitoring
Near-term revenue and validation path for motors, thermal plant, batteries, compute and rotating machinery.
From raw physical streams to trusted robot state
Physical signals
Force, torque, strain, vibration, motor current, heat, tactile fields and local impedance.
HTDF body-region chip
Multi-timescale analog state, baseline drift, coincidence, tokenisation and anomaly response.
Body-state tokens
Trust, fatigue, contact, degradation, unfamiliarity, confidence, permission and escalation.
AI / controller
Act, derate, suspend learning, trigger local reflex, request verification or escalate to main compute.
Validation roadmap
The near-term objective is not another broad benchmark. It is customer and silicon validation for a V1 CMOS companion chiplet while V2 substrate mechanisms continue in simulation.
Software validation
Public benchmark evidence across mechanical, thermal, electrochemical, chemical and compute systems.
Capability validation
Spatial localisation, graceful degradation, label-free unfamiliarity detection and manufacturing-tolerance simulation.
Power-latency characterisation
Circuit-level V1 estimates with confidence bounds for robotics body-region deployment.
Buyer validation
Customer-discovery conversations with surgical robotics, industrial cobot and humanoid integration teams.
V2 mechanism validation
Integrated dynamic substrate validation if the current screening phase produces positive signal.
V1 first silicon path
CMOS companion chiplet route through mixed-signal implementation, MPW and partner-facing demonstrator.
Partner access
Seeking evaluation partners and semiconductor collaborators for V1 physical-state inference validation in robotics, surgical systems, cobots, prosthetics and defence autonomy.