Edge AI
Run bounded inference close to the machine on appropriate MCU, NPU, GPU or edge compute, reducing dependence on continuous cloud connectivity where the workload permits it.
Physical AI / Edge execution infrastructure
VALO Physical
A physical execution layer for machines: local intelligence, sensing and consequence control connected through a governed path from observation to real-world effect.
Intelligence can propose. VALO controls what reaches the physical world.
01Edge AI hardware
02Local models
03Sensors & perception
04VALO Physical
05Machine / actuator
06Consequence + evidence
From sensing to consequence.
The physical world is not an API. State can be stale, incomplete or unsafe. The action path therefore has to survive uncertainty.
ObserveLocal sensors and models establish what is visible now — and what remains unknown.
ProposeIntelligence produces an intent. Intent is not permission.
GovernCurrent authority and constraints are checked before the effect path opens.
ActThe machine or actuator receives only the governed consequence.
VerifyObserved physical feedback is bound back to the action evidence.
One physical track. Four bounded capabilities.
The hardware, perception model and actuator can vary. The important invariant is the governed boundary before physical consequence and the evidence after it.
Machine Governance
Place a fresh authority and constraint boundary between machine intent and physical effect. A planner or model does not acquire permission merely because it produced an action.
Retrofit Intelligence
Add sensing, perception and governed action to useful existing equipment instead of requiring wholesale replacement of the installed base.
Physical Evidence
Preserve what the machine observed, what remained unknown, what action was allowed or denied, and what physical feedback was actually observed after execution.
PHYSICAL EFFECT PATH
From sensor to receipt.
The compute class can change. The consequence path must remain explicit.
01Sensor
02Local inference
03Proposed physical consequence
04VALO Physical
05Fresh authorization
06Gateway
07Machine
08Veritas receipt
HARDWARE & REFERENCE TARGETS
Designed across the physical compute ladder.
These rows describe integration scope, not blanket hardware validation. This web repository contains no device-specific validation evidence for STM32, Nordic, NPU/SoC or edge-GPU targets; they are therefore labelled target or research/evaluation rather than tested.
Implemented / testedIntegration targetResearch / evaluation
MCU
Integration target
STM32-class · Nordic-class
Bounded sensing, local preprocessing and control-adjacent integration where workload, memory and timing constraints permit. Device-specific performance and safety remain to be validated.
Edge SoC / NPU
Integration target
NPU / SoC class
Local perception and inference close to the machine, with VALO Physical governing the transition from proposed consequence to authorized effect.
Edge GPU
Research / evaluation
GPU-class edge compute
Target for higher-throughput perception and multi-model workloads. No hardware-specific benchmark or interoperability claim is made here.
Industrial / Vehicle Compute
Research / evaluation
Industrial controllers · vehicle compute
Reference target for systems where physical effects cross safety, operational and regulatory boundaries. Validation is installation-specific.
REFERENCE IMPLEMENTATIONS / DEMONSTRATORS
Evidence before hardware claims.
Governed Effect Path
The established architecture requires externally consequential paths to cross current authorization, enforcement and evidence capture. This is implementation-level governance evidence, not proof for a named hardware target.
Inspect the architecture →Universal Retrofit Intelligence
Applied research for adding sensing, local intelligence and governed action to installed equipment. Hardware-specific validation remains target-dependent.
Open research track →MA reference architecture
MA is a reference architecture and proving ground for making physical environments digitally interactable. It is not a separate VALO product family and does not establish completed partner hardware integration.
Open MA →PHYSICAL INVARIANTS
Capability does not open the effect path.
NO_DIRECT_EFFECT_PATHNo ungoverned causal route to consequence.NO_DETECTION ≠ SAFEFailure to detect a hazard is not evidence of safety.Capability ≠ AuthorityA machine's ability to act never creates its right to act.VERTICAL PROFILES
One stack. Different physical environments.
These are deployment profiles over the same VALO Physical consequence path, not separate product stacks.
AutomotiveRoboticsIndustrial machineryMaritimeDronesField systems
Product above. Research underneath.
VALO Physical is the product surface. Universal Retrofit Intelligence remains the applied research track for retrofit architecture. Edge / MCU / NPU remains a technical substrate, not a standalone product.
Universal Retrofit IntelligenceApplied research · sensing, perception and governed action for existing infrastructure →Machine Action InfrastructureThe wider VALO action boundary from machine intent to authorized consequence and proof →Human Control ResearchResearch into meaningful intervention and control at consequential boundaries →
Physical AI does not mean uncontrolled autonomy.
Local sensing and inference can reduce latency, bandwidth and cloud dependence. They do not themselves grant a right to move a machine, open a valve, energize equipment or create another physical effect.
VALO Physical carries the Machine Action Infrastructure principle into the physical world: intelligence proposes; current authority and constraints govern consequence; evidence records the result.
Status: product architecture assembled from active VALO execution-governance work and applied edge/retrofit research. Hardware-specific performance, safety, interoperability and regulatory claims require validation on the selected device and deployment.