<aside> 🧱

What this is

A reusable reference for building a robot / physical AI perception stack with stable interfaces, measurable performance, and graceful degradation.

What this is not

• Can incorporate additional sensors (LiDAR, radar) later without breaking upstream interfaces.

A vehicle-specific implementation doc. Vehicle, sensor, and compute choices belong in the “Platform profile” section.

</aside>

1) Goal

Deliver a camera-first perception pipeline that:

• Runs on an edge compute target.
• Publishes a stable interface contract to downstream consumers (planning, autonomy, UI).

<aside> ✅

Definition of Done (overall)

• Perception publishes a stable contract topic set (messages, frames, rates).
• Latency and FPS are measured and logged in real time.
• Pipeline runs for sustained duration without crashes and with bounded degradation behavior. </aside>

2) Platform profile (fill per deployment)

Use this section to specialize the stack for a particular robot / vehicle.

• Middleware: ROS 2
• Compute target: (e.g., Jetson Orin Nano, x86 + GPU, etc.)
• Primary sensor: (e.g., front camera)
• Secondary sensors (optional): (e.g., LiDAR later)
• Operating modes:
  • camera-only (default)
  • fused (when secondary sensors are present)
• Out of scope: HiL (unless explicitly added)

3) Interface contract (freeze early)

These are the outputs that downstream modules will consume. Keep names and semantics stable across platforms.

Required outputs

• /perception/drivable_grid
  • Type: choose one and freeze: nav_msgs/OccupancyGrid or grid_map-style grid
  • Frame: choose one and freeze: base_link or odom
  • Update rate target: 10 to 20 Hz