Autonomy & Robotics
Autonomy and Robotics Platform
A platform that lets robots and drones process sensor data on the device to move and work on their own.
It covers autonomous patrols by mobile robots and drones, manipulation learning for robot arms and autonomous port crane handling.
Field challenges
Robot platforms differ in their sensors, working environments and safety constraints. Patrol routes and task targets also change from site to site, so fixed rules alone cannot automate the work.
How it works
Robots and drones navigate with prebuilt 3D maps and SLAM, and alert the operator when they detect an anomaly. The manipulation policies of robot arms are learned by fine-tuning VLA foundation models on demonstration data.
Where it applies
- Wheeled mobile robots
- Drones
- Quadrupeds
- Mobile manipulators
- Cobots
- Port cranes
System
On-device AI processes data from patrol robots, drones and fixed CCTV on site and sends only anomalies to the operator.
- Patrol robot
- LiDAR and camera
- Drone
- Fixed CCTV
01
Autonomous mobile robots
Wheeled mobile robots and drones patrol set routes on their own and detect anomalies such as fire, smoke and falls. We are applying this to market, library and forest services in a 2026 on-device AI public service pilot.
- Indoor navigation with prebuilt 3D maps and SLAM
- Obstacle avoidance and automatic return to base
- On-device anomaly detection and operator alerts
- Drone patrols that send the coordinates of anomalies
02
VLA-based manipulation learning
Robot arms learn manipulation policies by fine-tuning VLA foundation models on demonstration data collected with a teleoperation rig. We have applied this to quadrupeds, mobile manipulators, cobots and stationary platforms.
- Demonstration data collection by teleoperation
- VLA foundation model fine-tuning
- Sensor-to-action end-to-end policy learning
- Sensor-fusion state estimation and motion planning
- Teleoperation rig
- Gripper
- Camera
- Mobile base
03
Autonomous port crane handling
To automate loading and unloading by STS and gantry cranes, we are developing technology that collects training data in a port digital twin and trains VLA models. Unstructured tasks that people do today, such as handling twistlock cones, are in scope.
- Training data collection in a port digital twin
- VLA-based crane motion learning
- Automatic twistlock cone handling
- STS crane
- Spreader
- Twistlock cone
- Container
- Terminal truck
From perception to learning
Perceive
Video, LiDAR, sound and location data describe the surroundings and the task target.
Decide
On-device AI decides on anomalies and the next action.
Act
Navigation, manipulation and operator alerts respond on site.
Learn
Sensor logs and demonstration data retrain the models.
Key capabilities
- Navigation with prebuilt 3D maps and SLAM
- Real-time obstacle-avoidance path planning
- On-device anomaly detection
- Operator alerts and control-room integration
- Demonstration-based manipulation learning
- VLA foundation model fine-tuning
- Real-time state estimation using sensor fusion
- Autonomous behavior under safety constraints
Deployment goals
- Autonomous monitoring of areas that are hard to patrol around the clock
- Operator alerts when an anomaly occurs
- Automation of unstructured manipulation tasks
Frequently asked questions
- Which robots does it apply to?
- We have applied it to quadrupeds, mobile manipulators, cobots and stationary platforms, and we are applying it to wheeled mobile robots and drones in a public service pilot. For a new platform, we check its sensors and control interface before setting the scope.
- Do you work with VLA models?
- Yes. We fine-tune VLA foundation models and train sensor-to-action end-to-end policies.
- Does it work where the connection drops?
- The on-device AI in the robots and drones decides on anomalies. Events that occur while the connection is down are designed to be stored on the device and sent in order once it reconnects.
- How do I get a quote?
- Tell us the site, the type of robot platform, the tasks to perform and which existing control systems it should connect to. We will propose a configuration and quote.