Autonomous Robotics & ROS 2 Architecture
Admiral provides an immutable, high-reliability edge operating system and orchestration platform engineered specifically for Autonomous Mobile Robots (AMRs), Automated Guided Vehicles (AGVs), unmanned submersibles (AUVs/ROVs), robotic manipulators, and physical AI perception nodes.
- Admiral Cloud Control Plane (app.admrl.co)
▼ Bi-directional Management, Telemetry & OTA Updates
- The Admiral Mesh (Resilient Transport Fabric)
▼ Encrypted State Synchronization & Workload Control
- Robot Compute Unit (Admiral OS & Daemonless Manager)
1. Zero-Overhead Direct Namespace Orchestration
Admiral orchestrates Linux namespaces and cgroups v2 directly without the overhead of a Docker daemon or containerd:
- Sub-Second Initialization: Process containers initialize directly at the kernel boundary, providing near-instant startup and recovery.
- Minimal Memory Footprint: The entire Admiral manager and root OS require only ~100MB of RAM (hardware dependent), and achieve sub-10-second cold boots to workload ready on fast hardware, reserving physical memory for memory-intensive localization (SLAM), high-resolution voxel grids, and neural perception models.
- Deterministic Single-Workload Model: Admiral runs one primary workload container per machine, matching standard robotics practices where the entire on-robot software stack is orchestrated via a master launch file (
ros2 launch robot_bringup bringup.launch.py) or composable node containers (rclcpp_components).
2. Unrestricted DDS Multicast (hostNetwork: true)
Standard container runtimes isolate network namespaces with NAT bridges, which break ROS 2 DDS (Data Distribution Service) peer discovery.
In Admiral, toggling Host Network binds the workload directly to the machine's physical network adapters (eth0, wlan0):
- Native Multicast Discovery: UDP multicast discovery packets (
239.255.0.1) broadcast freely across local robot subnets. - Multi-Computer Subsystems: Easily link primary x86 navigation computers to secondary Nvidia Jetson perception nodes or auxiliary sensor microcontrollers without complex proxy routing.
- Off-Board Teleoperation: Developer workstations running RViz2, PlotJuggler, or Foxglove Studio on the same local network automatically detect ROS 2 topics when using matching
ROS_DOMAIN_IDsettings.
3. Real-Time Determinism & CAP_SYS_NICE
Robotics control loops (such as ros2_control, differential drive controllers, and inverted pendulum balancers) require strict timing guarantees.
Admiral grants CAP_SYS_NICE and CAP_SYS_RESOURCE to containerized workloads:
- Nodes can set real-time thread priorities (
SCHED_FIFOorSCHED_RR). - Kernel scheduler jitter is eliminated, preventing missed sensor sampling windows or dropped actuator control cycles.
4. Zero-Copy Shared Memory IPC (/dev/shm)
Stereoscopic cameras, high-rate IMUs, and 3D LiDAR point clouds generate high data throughput. Transferring serialized messages over standard network sockets strains CPU cores with memory copies.
ROS 2 supports zero-copy loaned messages (via FastDDS SHM or Iceoryx) using POSIX shared memory:
- In Admiral, you can allocate a sized
tmpfsmount (e.g.2GBto8GB) on/dev/shmwithin your configuration. - Camera and LiDAR drivers write raw sensor buffers directly into shared memory, allowing SLAM and perception nodes to access the data with zero serialization overhead.
5. Full Hardware & Sensor Pass-Through
With Full Device Access enabled, Admiral mounts the entire host /dev tree into the workload, granting direct access to robot peripherals:
| Subsystem | Linux Node | Consumption in ROS / ROS 2 |
|---|---|---|
| Motor Controllers / IMUs | /dev/ttyUSB*, /dev/ttyACM* | Serial drivers, micro-ros-agent, or Roboteq/Odrive motor controller nodes. |
| CAN Bus Actuators | can0, can1 (SocketCAN) | Industrial servos, steer-by-wire steering columns, and smart BMS batteries via standard Linux SocketCAN. |
| RGB-D Cameras & LiDAR | /dev/video*, /dev/bus/usb | Intel RealSense, Luxonis OAK-D, and USB LiDARs upload firmware and stream point clouds directly. |
| Physical E-Stop & Bumpers | /dev/input/event* | Hardware emergency stop buttons and bumper microswitches feed into safety supervisor nodes. |
| AI Perception Accelerators | /dev/nvidia*, /dev/rknpu | Automatic driver mapping enables Nvidia CUDA/TensorRT and Rockchip NPU hardware acceleration. |
6. Mission-Aware Canary Rollouts
Deploying software updates to autonomous machines while they are actively navigating warehouse floors or public spaces presents safety and operational hazards.
Admiral provides controlled deployment workflows:
- Docked Update Windows: Schedule configuration rollouts to occur only during designated charging or maintenance windows.
- Canary Phasing: Deploy navigation or perception model updates to a single robot canary batch (
1–5%of fleet) before promoting fleet-wide. - Automated Health Rollbacks: If the updated workload fails health checks or loses connectivity with the Admiral Mesh, the rollback engine reverts the robot to the previous known-stable configuration before it departs on its next mission.
7. Storage Protection with Copy-on-Write Subvolumes
Robotic diagnostic bags (ros2 bag record) and persistent navigation maps are stored on dedicated Copy-on-Write (CoW) subvolumes:
- Persistent Subvolume Isolation: Separate volatile telemetry and bag recording directories from persistent configuration and application state.
- Instant Map Checkpoints: Point-cloud costmaps and topological navigation graphs saved to subvolumes benefit from instantaneous atomic snapshots.
- Power-Cut Resilience: The underlying CoW storage engine guarantees filesystem integrity during unexpected battery disconnects or emergency stops.