Open spatial
intelligence

The modern framework for robot perception.

Get started→Get started→Read the docs

BACKED BY

THE STACK

Three products.
One robot brain.

Three products that answer three different questions. Use one, or all three — they share one interface and one world.

NOW

Perception

See what’s around the robot right now — objects, people, free space, depth — from any camera or lidar.

Explore perception →

WHAT

Scene understanding

Understand what the place is — objects, surfaces, rooms and how they relate — as a live 3D scene graph the robot can query.

Explore scene understanding →

World models

Predict what happens next — how people, doors and carts will move — so the robot can plan, not just react.

Explore world models →

Perception
See what’s around the robot right now — from any camera or lidar.
Scene understanding
Understand what the place is — a live 3D scene graph of objects, surfaces and rooms.
World models
Predict what happens next, so the robot can plan — not just react.

SDK

One Python SDK for the whole stack.

Dead simple. Pick an algorithm, feed it frames, read the result — pose, map, landmarks, loop closures — in the same loop.

pip install opensense: Python 3.10+, Linux, CUDA optional
Typed, async, hot-swappable: Every layer is a protocol

SDK reference→

main.py

from opensense import Camera, SLAM

cam = Camera.realsense()

slam = SLAM(device="orin")

for frame in cam.stream():

out = slam(frame)

print(out.pose, out.map.size_m2, out.landmarks, out.loop_closures)

slam.map.save("warehouse-a")

from opensense import Camera, SLAM

cam = Camera.realsense()

slam = SLAM(device="orin")

for frame in cam.stream():

out = slam(frame)

print(out.pose, out.landmarks)

opensense

warehouse-quad · Vektor Robotics

Good afternoon, Ada.

Here’s what your robot has seen and how well your models see it.

Upload dataset New evaluation

Upload a datasetDrop a rosbag or a folder of frames.→

Run an evaluationScore a model on any dataset.→

Deploy a modelPush a checkpoint to the robot.→

What the robot saw One scene, four runs · warehouse-a · 41k frames

All datasets →

pallet

person

forklift

warehouse-a / run-01

day · 12k frames · cam ×2 · lidar

Ready

dock

cone

warehouse-a / run-02

night · 14k frames · cam ×2 · lidar

Ready

warehouse-a / run-03

forklift traffic · 9k frames

Processing

person

warehouse-a / run-04

stacked pallets · 6k frames

Needs labels

How well the models see Latest evaluation · det-1.5 vs det-1.4 on warehouse-a / run-02 (night)

All evaluations →

mAP@0.5

0.74

+0.03 vs det-1.4

Recall · person

0.91

+0.06 vs det-1.4

Latency · Orin NX

23 ms

+2 ms vs det-1.4

person

0.91

+0.06

pallet

0.88

+0.01

forklift

0.79

+0.03

cone

0.72

+0.02

door

0.61

+0.09

cart

0.55

-0.03

Regression: cart −0.03 · 12 failure frames queued for labeling

Where it got it wrong

12 frames →

missed person

run-02 · low light

pallet → cart

run-04 · pallet vs cart

ghost person

run-01 · floor reflection

missed cone

run-03 · occluded by forklift

CONSOLE

Run the fleet from one place.

Datasets, evaluations, API keys and usage in a hosted console. Free for one robot; plans that scale with the fleet.

Every robot, scene and model in a hosted console: datasets, evaluations, latency, and one-click model deployments across sites. Free for one robot; plans that scale with the fleet.

Free1 robot · local scenes
Teamup to 25 robots · hosted scenes · alerts
Fleetunlimited · multi-site · SLA

See plans→See plans→

1 min

to first scene graph

40+

sensor drivers

robot bodies supported

cloud dependencies

Give your robot
a horizon.

Free to start. Hosted plans when your fleet grows.

Get started→

Get started→Read the docs

Open spatialintelligence