CORE CONCEPTS

Point clouds

Point clouds describe physical objects and spaces as 3D coordinates. Group points by object id and kind so Rowe can reason about fit, placement, stability, motion, and scene structure.

Group points by object

A point cloud with no object ids is a shape, not a scene. Rowe reasons about relationships between things, so every group you want to ask a question about needs its own object_id. Tasks like fit and placement address objects by that id.

object_id — how you refer to the group in a task
kind — box, container, surface, or another hint about what it is
points — the coordinates themselves, in one consistent unit

Units are yours to keep straight

Coordinates are taken as given. There is no unit inference and no normalization, so a scene mixing millimetres and metres will produce confident, wrong measurements. Convert before you send.

When to use CAD instead

Point clouds are the right input for scanned or sampled geometry. If you have the CAD file, send a cad_scene — it keeps exact volumes, faces, and dimensions that a sampled cloud can only approximate.

Request fields

The full input contract, shared by every task.

modelstring

Model id. Use rowe-1.0.

input.task_typestring

fit, stability, placement, embed, understand, reconstruct, predict, act, or optimize.

input.cad_sceneobject

Native CAD scene with B-rep solids, drawing entities, dimensions, constraints, and optional feature_tree. Preferred for STEP, IGES, DXF, PDF, SVG, and draw.io workflows.

input.previous_cad_sceneobject

Optional previous native CAD scene for compare and temporal tasks.

input.point_cloud_sceneobject

Current 3D scene with one or more point cloud groups.

input.previous_point_cloud_sceneobject

Optional previous scene for temporal prediction.

input.image / input.imagesobject

Optional raster frames, depth maps, or multi-view visual context.

input.sensor_readingsobject

Optional IMU, GPS, BLE, WiFi, altimeter, or device state readings.

input.goalobject

Optional target state for action, routing, placement, or optimization tasks.

input.object_idstring

Primary object id used by fit, stability, placement, and tracking tasks.

input.secondary_idstring

Target or container object id for fit and placement tasks.

input.detailstring

Use compact to reduce structured prediction detail for small or latency-sensitive requests.