Physical AI needs a plant it can read.
A robot arm, an AMR, an inspection drone, a vision cell — each one perceives its own few metres and nothing beyond them. None of them knows what that asset is called in SAP, which zone is under a permit to work, what the drawing says the valve should do, or who is allowed to authorise the next move. PredCo builds the layer that answers those questions, on the cameras, sensors, and edge hardware already in your plant.
- PerceiveExisting cameras
- GroundResolved assets
- DecidePlant context
- ActScoped tools
- GovernFull audit trail
Physical AI does not stall on the robot. It stalls on the plant.
The hardware works. What is missing on almost every floor we walk is a record of the plant accurate enough for a machine to act against, and governed enough for anyone to sign off on it.
Perceive, ground, decide, act, govern.
Five stages between a signal on the floor and an action a plant will authorise. Each one is assembled from modules that already run in production, which is why this is an extension of the Data Core rather than a new product.
- Perceive01
The floor becomes a signal, not a video feed
Vision models run on the cameras already mounted above the line, alongside PLC, historian, and sensor telemetry. The output is a structured event with a timestamp, a location, and an asset — not a frame for someone to watch.
Vision EngineConnector HubEdge Runtime - Ground02
Every event is tied to a real, named thing
Entity resolution reconciles the cell in the PLC, the asset in SAP, the tag in SCADA, and the object in a robot's own map into one record. Without this step an autonomous machine is acting on a guess about what it is looking at.
Entity ResolutionKnowledge BaseDocument Intelligence - Decide03
Context the machine cannot sense for itself
Batch, shift, work order, permit status, maintenance history, drawing topology, and regulation are all in the same record. The decision is made against the state of the plant, not against the state of one camera frame.
Contextual EnginePredictive EngineAnomaly Detection - Act04
An instruction with an owner and a boundary
Agents plan and act through tools scoped by the same role permissions that apply to people. The output is a work order, a dispatch, a hold, a setpoint change, or an escalation — issued into the system that already governs it.
Agent StudioWorkflow BuilderAlerting & Routing - Govern05
The reason the plant is allowed to switch it on
Every plan, call, and result is written to the audit trail with full lineage back to the source reading. Confidence below threshold escalates to a named human. This is the part that decides whether autonomy clears EHS, not the model.
Lineage & Audit TrailRole-Based Access ControlHuman-in-the-Loop Review
The same five stages, on a real incident.
A camera detection on its own is seven fields and no context. Step through what the Data Core joins onto it, and where each fact comes from — the growth of this record is the entire argument for the layer.
A person enters a press cell while the line is running.
Stamping · line STAMP-L2 · camera CAM-38-04 · 14:02 on B shift
Perceive. The edge appliance turns a frame into a typed event. This is everything a camera or a robot controller can know by itself, and it is not yet enough to act on.
- event.id
- evt_7f3a91c4
- Edge Runtime
- event.ts
- 2026-09-03T14:02:11.324+05:30
- Edge Runtime
- source.camera
- CAM-38-04 · RTSP
- Existing CCTV
- detect.class
- person
- Vision Engine
- detect.confidence
- 0.94
- Vision Engine
- detect.bbox
- [412, 288, 96, 214]
- Vision Engine
- infer.latency_ms
- 312
- On-appliance
- asset.id
- PRESS-104
- Golden record
- asset.sap
- EQ-338104
- SAP PM
- asset.cmms
- PR-104
- Maximo
- asset.scada_tag
- U38_PRS_04
- SCADA
- zone.id
- Z-PRESS-3 · exclusion
- Geofence
- line.id
- STAMP-L2
- Plant ontology
- batch.id
- B-88213
- MES
- shift
- B
- Roster
- machine.state
- RUNNING
- OPC-UA
- machine.stroke_rate
- 28 spm
- OPC-UA
- permit.active
- none
- PTW register
- policy.matched
- EXCL_ZONE_WHILE_RUNNING
- Rulebook v14
- severity
- high
- Contextual Engine
- precedent.30d
- 3 similar · same zone
- Event history
- action.1
- notify → supervisor, shift B
- tool: alert_route
- action.2
- hold.advisory → STAMP-L2 HMI
- tool: line_advisory
- action.3
- incident.open → EHS register
- tool: ehs_write
- acting_as
- role EHS_SUPERVISOR
- RBAC
- escalation
- 0.94 ≥ 0.90 threshold → auto
- Threshold policy
- elapsed_ms
- 1,420 · detection to alert
- Measured
- model.version
- vision-ppe-4.2.1 · pinned
- Model Registry
- policy.version
- rulebook v14 · eff. 2026-07-01
- Version Control
- decided_by
- agent:zone-guard
- Agent Studio
- lineage
- frame → detect → asset → policy → action
- Lineage
- frame.egress
- none · never left site
- Edge Runtime
- audit.ref
- aud_2f81ba09 · immutable
- Audit Trail
What runs today, what we build with you, and what we do not build at all.
Physical AI attracts more claims than deliveries. This is the line we hold in a technical evaluation, so you can hold us to it.
Deployed across 25+ plants. This is product, not roadmap.
- Vision models on existing plant cameras — PPE, exclusion zones, cycle time, foreign body, surface defect
- Edge inference on NVIDIA Jetson / IGX or your own GPU, air-gap capable
- Events grounded to a resolved asset, line, batch, and shift
- Predictive and anomaly models on machine, thermal, vibration, and current signals
- Agentic workflows that plan and act inside scoped permissions, with every step audited
Scoped per site, delivered as a joint engineering programme.
- Actuation into machine and robot controllers over OPC-UA, MQTT, or a ROS 2 bridge
- AMR and fleet dispatch driven by plant context rather than a standalone fleet manager
- Closed-loop setpoint control where the process and the safety case allow it
- Robot-cell quality feedback — inspection result returned to the cell in-cycle
- A queryable spatial and topological model of the site from drawings, scans, and camera geometry
We do not build hardware, and we will not quote you as if we do.
- Robot arms, cobots, AMRs, AGVs, and inspection drones
- Safety-rated controllers, light curtains, scanners, and E-stop architecture
- Mechanical integration, guarding, and cell commissioning
- Machine builders and robotics integrators we work alongside on delivery
The places autonomy pays before it impresses.
Every one of these starts with perception and grounding on equipment you already own, and only then extends to a machine that acts.
Functional safety stays where it belongs: in the safety-rated controller, the interlocks, and the guarding designed for that machine. PredCo sits above that layer. What we supply is the evidence, context, and audit trail that a risk assessment, an EHS committee, and an auditor need before an autonomous machine is allowed to run.
- No PredCo output is on the safety-critical path of a machine
- Every autonomous action is attributable to a policy, a model version, and a person
- Confidence thresholds escalate to a named human rather than proceeding
- Model versions are pinned, tracked, and rollback-capable per site
- Deployment is on-premise or air-gapped, so nothing leaves the perimeter to be inferred on
PredCo does not certify a machine against these. We produce the evidence and traceability the assessment, the notified body, and the auditor ask for.
- Existing RTSP / ONVIF
- Thermal · vibration · current
- PLC tags & historian
- Robot cells · cobots
- AMRs · AGVs
- Presses · conveyors · utilities
- Edge Runtime
- Vision
- Grounding
- Context
- Agents
- Audit
Reads the signals, resolves them to the plant record, decides against context the machine cannot sense, acts through scoped tools, and writes every step to the audit trail.
- Interlocks · light curtains
- E-stop · safety PLC
- Certified to ISO 13849 / IEC 62061
Independent of PredCo. Nothing we produce sits on this path.
- Supervisor alert
- Work order
- Advisory to HMI
- Audit record
What your controls and IT teams will ask first.
The integration surface, stated before the workshop rather than discovered during it.
- Machine & control
- OPC-UA · Modbus TCP · MQTT · Sparkplug B · PROFINET gateways
- Robotics
- ROS 2 bridge · vendor controller APIs · fleet-manager APIs
- Perception input
- Any RTSP / ONVIF camera · thermal · lidar · vibration · current
- Edge compute
- NVIDIA Jetson · IGX · customer-supplied GPU appliance
- Systems of record
- SAP · Oracle · Maximo · CMMS · MES · historian
- Identity & access
- SAML 2.0 · OIDC · Azure AD · Okta · role-based
- Deployment
- On-premise · Air-gapped · Private cloud · Hybrid
- Egress
- Optional. Raw video and telemetry need never leave the site.
The five questions every enquiry opens with.
What does PredCo mean by Physical AI?
AI that perceives and acts in the physical world rather than only on a screen — robots, AMRs, drones, vision cells, and autonomous machinery on a plant floor. PredCo's role in it is specific: we are the perception, grounding, and governance layer those machines act against. We build the plant record that makes an autonomous action correct and auditable. We do not build the hardware.
Does PredCo build robots?
No. Robot arms, cobots, AMRs, drones, and safety-rated controllers come from machine builders and robotics integrators, several of whom we deliver alongside. PredCo supplies the layer above the hardware: perception on existing cameras, a resolved record of every asset and zone, the decision context, and the audit trail. If a supplier tells you they do both the fleet and the intelligence layer equally well, ask which one they actually ship.
Is PredCo a safety system?
No, and this matters. Functional safety stays in the safety-rated controller, the interlocks, and the guarding for that machine. Nothing PredCo produces sits on a machine's safety-critical path. What we supply is the evidence layer above it — what happened, on which asset, under which policy, decided by which model version, escalated to which person.
Can this run without sending video or telemetry off site?
Yes. Inference runs on an edge appliance inside the plant — NVIDIA Jetson, IGX, or your own GPU — and only the structured event leaves the box. Fully air-gapped deployment is supported, which is how we work with PSUs and critical infrastructure operators.
Where should a plant start with Physical AI?
Almost never with a robot. Start with perception and grounding on the cameras and signals already installed, because that is what every later autonomous step reads from — and it pays for itself on safety, quality, and uptime before a single machine is given authority to act. Plants that buy the robot first usually spend the following year building this layer anyway, under schedule pressure.
Bring one problem. We'll show you the deployment.
A 30-minute working session with our engineering team. No slideware. Bring a real operational problem and leave with a concrete view of what a PredCo deployment looks like for your plant.