Built for accountable physical AI

Behavior contracts

Reusable, versioned behaviors — patrol, dock, inspect, timed drives — with safety rules and preconditions built in. Contracts are hashed, registered, and referenced on every intent so you always know exactly which recipe ran.

Not one-off prompts. Not ad-hoc scripts at runtime. Behaviors are reviewed, tested, and promoted like software releases.

• YAML-defined contracts with schema validation
• Preconditions: battery, localization, clearance gates
• Invariants enforced continuously during execution
• Version pinning via contract reference on each intent

Intent interface

Any AI agent, operator tool, or integration submits structured intents — goals and parameters, not motor commands. Groundpasse accepts the request, validates it, schedules execution, and exposes status over HTTP and JSON-RPC.

Agents never publish directly to cmd_vel or joint topics. They describe outcomes; the runtime owns motion.

• REST intent API with polling and cancellation
• Python and TypeScript client adapters
• Deadline and priority on submitted jobs
• Shadow mode for parallel sim comparison

Safety governor

A dedicated safety layer sits on every motion command path. Speed caps, proximity stops, zone violations — clamped or blocked before actuators respond. Agents and contracts cannot bypass this layer; only physical e-stop or operator override can.

That is the ground pass: explicit safety approval before anything touches ground.

• Per-command velocity and acceleration limits
• Intervention logging inside every episode
• ROS 2 bridge applies governor on live hardware
• Independent of agent reasoning — deterministic enforcement

Episode records

Every run produces a complete audit trail: what was requested, which contract ran, every precondition result, governor intervention, sensor snapshot, and final outcome. Episodes are the diary of physical AI — replayable, diffable, certifiable.

• Web viewer for timeline and event inspection
• Diff two episodes to spot behavioral regression
• Certification reports for compliance workflows
• Retention policies for long-running deployments

Behavior benchmarks

Standard scenarios score how reliably behaviors perform — in simulation or on real hardware. Run benchmarks in CI, compare across contract versions, and catch regressions before they reach the floor.

• Curated scenario library (patrol, dock, inspect, …)
• Sim and real execution paths
• Automatable via make bench in development
• Marketplace catalog for shared contract templates