Perception + Planning Stack

The Mobvynt Perception & Planning Stack.

A software layer that runs on your robot's compute, takes sensor inputs, and returns collision-free paths — with obstacle type classification included.

Modules

Three modules. One coherent runtime.

MV Perceive
Sensor fusion + classification

Ingests LiDAR point clouds and optional camera frames. Runs background subtraction, moving object isolation, and an 8-class CNN classifier to identify person, forklift, pallet, conveyor, static shelving, mobile cart, unknown-moving, and unknown-static. Trajectory prediction using a constant-velocity linear model gives the planner 2–3 cycles of lookahead.

Detection latency
< 80ms
Obstacle classes
8 standard + safe defaults
Sensor input
2D LiDAR minimum
Fusion modes
LiDAR-only, RGBD+LiDAR
MV Plan
Dynamic path re-planning

Graph-based path search using a modified A* that threads two constraints simultaneously: obstacle clearance from MV Perceive's obstacle map, and pick-slot arrival window from your WMS. Configurable risk tolerance parameter lets you tune the trade-off between path safety margin and deadline adherence.

Re-plan cycle
< 120ms
Path success rate
97% live deployments
Algorithm
Modified A* + temporal constraints
Output rate
20Hz velocity commands
MV Bridge
Integration & fleet API

ROS 2 integration layer that handles hardware abstraction and fleet management system connectivity. Publishes to standard ROS 2 Navigation2 topics so no firmware changes are needed. REST or MQTT adapter connects to your WMS and fleet management system for timing window ingestion and fleet-level obstacle event sharing.

ROS 2 versions
Humble, Iron, Jazzy
Fleet systems
12+ tested configurations
WMS protocols
REST, MQTT, file polling
Firmware changes
None required
Architecture

Runs on-robot. No cloud dependency.

Standard ROS 2 message interfaces so it drops into existing stacks. Configurable sensor inputs — works with 2D LiDAR-only, 2D+3D, or camera+LiDAR fusion.

2D LiDAR /scan RGBD Camera /camera/image MV Perceive sensor fusion object classify traj predict /mobvynt/obstacles Obstacle Map temporal costmap decay functions MV Plan A* + timing constraint solver risk tolerance config /cmd_vel /mobvynt/path Motor Ctrl 20Hz vel cmds no fw changes MV Bridge — ROS 2 integration / WMS timing window input / Fleet API REST / MQTT / file polling — no firmware changes required
Performance

Numbers from live deployments.

< 80ms
Detection latency
End-to-end from sensor frame arrival to obstacle classification output. Measured across 2D LiDAR-only configuration on Intel NUC-class hardware.
< 120ms
Re-plan cycle time
From obstacle detection trigger to new path output published on /mobvynt/path. Includes costmap update and A* re-solve with timing constraints.
97%
Path success rate
Measured across logged deployment sessions in ambient fulfillment environments. A path is considered successful when the robot reaches its pick-slot destination without operator intervention.
8
Obstacle classes
Pre-trained: person, forklift, pallet, conveyor, static shelving, mobile cart, unknown-moving, unknown-static. Unknown classes use safe conservative defaults rather than failing.
7
Sensor configurations
Tested hardware configurations ranging from single 2D LiDAR to 2D+3D LiDAR fusion to camera+LiDAR. Each configuration has its own tuned inference profile.

Ready to drop it in your stack?

Tell us your robot model and ROS 2 version. We'll walk you through the integration and show you what the numbers look like for your specific setup.

Request Access Schedule a Technical Call