Spatial AI · Perception · Autonomy

A machine can only act on space it can measure.

spacejump builds the perception stack that turns raw lidar, radar and camera into a metric map — geometry, free space, and everything that moves. It runs on the robot, at 40 Hz, with no connection to anywhere.

SENSOR
VIEW
LIVE SCAN RETURNS 6 400 ROTATION 20.0 Hz TRACKED 3 DRIFT 0.04 m / km LATENCY 12.7 ms

Four stages between a photon and a wheel command.

Each stage owns a latency budget. Nothing downstream is allowed to wait on the cloud, and nothing upstream is allowed to guess.

+0.8 msINGEST

Raw returnslidar · radar · stereo · IMU

Frames arrive time-aligned and motion-compensated. A spinning sensor on a moving base skews every sweep it takes — we undo that against wheel odometry and IMU before a single point reaches the next stage.

+4.2 msGEOMETRY

Structureground plane · occupancy · mesh

Points become surfaces. We fit the drivable ground, carve free space out of the occupancy field, and grow a metric mesh that still closes its loops after eight hours of continuous operation.

+5.6 msSEMANTICS

Things that movedetection · tracking · velocity

Static structure separates from dynamic objects. Every track carries an identity, a velocity, and an honest covariance — a pedestrian half-occluded behind a pallet stack is reported as exactly that, not as empty road.

+2.1 msMOTION

A path to followcost map · plan · uncertainty

The controller gets a trajectory and the confidence attached to it. Where the map is thin, the plan slows down instead of pretending.

END TO END 12.7 ms ON A 40 W MODULE · NO NETWORK IN THE LOOP

Built for places that were never surveyed.

Structured environments are the easy case. Our customers work where the floor plan changed this morning.

CONSTRUCTION

Sites that change daily

Reconstruct a building from a walkthrough, then keep the model current as the structure goes up. Measurements hold to centimetres without survey markers.

INTRALOGISTICS

Warehouses with people in them

AMRs that read a shared aisle correctly — including the forklift coming the other way and the pallet that was not there on the last pass.

INSPECTION

Structures you can't touch

Drones and crawlers holding six-degree-of-freedom pose against steel and concrete, with no GPS and no prior map.

OFF-HIGHWAY

Ground without lane markings

Ports, quarries, and yards, where the drivable surface has to be inferred from geometry rather than read off paint.

A parking space is a measurement, not a painted rectangle.

The same four stages at road speed. Surfaces accumulate as the car passes, every gap is measured against the car's own footprint, and the first one that actually fits becomes a trajectory.

  • Occupied
  • Gap too narrow
  • Obstructed
  • Free · fits

MINIMUM GAP 2.35 m = VEHICLE WIDTH + DOOR CLEARANCE · GAP MEASURED TO ±0.04 m · REPLANNED AT 12 Hz

Everything a perception team asks in the first call.

Answered here, so the call can be about your sensors instead.

SensorsMechanical and solid-state lidar, stereo and mono camera, 4D radar, IMU, wheel odometry
ComputeJetson Orin NX · Jetson Thor · x86 with a discrete GPU
Throughput2.4 M points / s sustained, 40 Hz output
Loop closure drift0.04 m / km on closed trajectories
InterfacesROS 2 · MCAP and Foxglove · gRPC · C++17 headers
CalibrationTargetless extrinsics, re-solved online
NetworkNone required. Maps stay on the machine.

“We had six months of logs and no ground truth. spacejump gave us a map that closed, and told us where it didn't trust itself. That second part is the reason we shipped.”

Head of Autonomy · off-highway OEM

The hard part isn't seeing the litter. It's knowing what to leave.

A street cleaner makes one U-shaped pass across the plaza: out along the kerb, up, and back along the frontage. Everything it finds is classified before the arm moves, because a drain grate and a crushed can look much the same from two metres up.

  • Litter · collect
  • Fixed · ignore
  • Organic · leave
  • Unclassified

SCAN RADIUS 5.5 m · CLASSIFY BEFORE GRASP · BIN 4 ITEMS · REJECTS ARE LOGGED, NOT DISCARDED

Start with your own data

Send us 10 minutes of logs.

We'll run our stack over your recordings and come back with a map, the tracks, and an honest list of what your sensor setup can't currently see.