Physical AI
The authority layer for robots, cobots, AGVs, drones, and autonomous machines
When agentic systems are about to take a real-world action, Intended classifies the intent, mints a short-lived signed credential bound to the robot's identity, and falls back to a declared safe-default if the credential is denied or expires.
Layered on top of your existing safety PLC — never in place of it. Same shape as Snyk for code or Cloudflare for traffic: we provide the primitives; you wire them into your robot.
Where we sit
Above your safety PLC, never in place of it.
The safety PLC is the authority on physical safety hazards. Intended is the authority on whether the command should be issued in the first place.
Layer 1
Operator HMI / autonomous planner / AI agent
Where intent originates
Layer 2
Intended layer (this is new)
Classify intent → mint short-lived Authority Token with sub-second TTL
Layer 3
Existing safety PLC + bus (unchanged)
E-stop chain, light curtains, presence scanners stay where they are
Layer 4
Motion controller → motors
Verifier checks the token at the edge, falls back to safe-default if denied
What you get, what you keep
Clear ownership boundaries.
We provide the certified primitives. You assemble them into your machine. Your roboticists, your controls engineers, your safety officer.
What we own
- OIL — open Apache 2.0 taxonomy, 173 categories across 29 domains
- Authority Runtime — cloud classifier, policy evaluator, signing issuer, audit chain
- Edge verifier — Rust crate, sub-50ms local verification, 24-hour offline operation
- SDKs — Python, TypeScript, ROS2 ament; C++ and Rust customer-facing SDKs ship with first design partner
- Reference cobot example — ROS2 / MoveIt-shaped pick-and-place that runs on a laptop
What you own
- Your robotics platform integration (ROS, ABB, FANUC, KUKA, UR, custom)
- Your safety bus wiring (FSoE, PROFIsafe, Pilz PSS, OPC-UA Safety)
- Your PLC ladder logic / SCL / function blocks — unchanged
- Your PhysicalStateProvider — bridges your bus into Intended predicates
- Your machine's safety case — we provide the certified primitive + GSN template
What ships today
Production-grade primitives, not hand-waving.
Structured-goal classification
Robotics agents emit ROS2 actions, OPC-UA method calls, MAVLink commands — not English. Intended classifies the structured payload into an OIL v2 category with a confidence score and safety bit. Fail-closed below the calibrated per-family floor.
Sub-second Authority Tokens
RS256 JWT bound to your robot's identity (IEEE 802.1AR DevID), the deadline declared on the DAG node, and the safe-default action. Token TTL bounded by the deadline — not minutes, not hours.
Edge verification
Rust crate with no_std core, multi-arch builds (amd64 / aarch64 / armv7), JWKS offline cache, revocation list sync, pre-computed authority caching. Hot-path verification target ≤50ms.
Continuous-time + consensus predicates
Policy DSL supports `if X for ≥250ms then deny`, `if 2 of 3 (human_sensor) then deny`, attestation requirements (FSoE / PROFIsafe / Pilz PSS), audit-gap policy.
Safety-rated input attestation
Predicates carry channel + protocol + safety-rated flag. Policy clauses can require safety-rated channels for SIL-relevant decisions; non-rated predicates can't satisfy them.
Tamper-evident audit chain
Every issuance, denial, escalation, and verification is recorded. PTP/NTP-attested per-event timestamps. Forensic replay against historical state snapshots.
Stack-native integration
Wherever your robot already lives.
Robotics platform (ROS2)
intended_ros2 ament package — classify_and_authorize service
Industrial controls (TwinCAT / Studio 5000 / TIA Portal)
Mediator pattern — non-RT process bridges PLC ↔ cloud
ML perception / planning (Python)
intended Python SDK — same shape as the reference cobot
Real-time motion controller (Rust)
intended-verifier crate — sub-50ms hot-path verification
Safety case authoring
Extend the certified primitive into your machine's safety case
Existing safety PLC
Stays exactly where it is — Intended layers above, never replaces
Verticals with shipped OIL coverage
Six physical / embodied verticals at v2.
The Open Intent Layer v2 spans 15 physical-AI domains. Your vertical's vocabulary is already in the taxonomy.
- Manufacturing — cobot pick-and-place, AGV / AMR mission steps, conveyor / sortation lines
- Surgical / medical robotics — primitive-level authorization, medication-dispensing audit
- Aerial — drone takeoff, waypoint navigation, payload release with geofence gating
- Autonomous vehicles — lane changes, mission planning, ASIL-B/D safety cases
- Energy / utilities — substation switching with LOTO, generation dispatch
- Mining + construction — drilling / blasting / lifting with exclusion-zone enforcement
What we do not claim
Honest scoping.
- We do not replace your safety PLC. The e-stop chain, light curtains, and safety-rated I/O stay exactly where they are.
- We are not currently SIL-2 / PLd certified. The verifier is buildable today; third-party assessment is sales-funded when a customer's regulatory posture demands it. Un-assessed safety-relevant SW is a defensible posture under your own SW lifecycle.
- We do not write your safety case. We provide the certified primitive + the GSN template; you author the assembly's case.
- We do not provide platform-specific adapters for ABB, FANUC, KUKA, UR, Boston Dynamics, etc. Partner / community work against the same SDKs.
Get started
Five minutes, no robot required.
Sign up, run the reference cobot against the live cloud, see a real Authority Token signed and decode it.