SLAM & Localisation: ORB-SLAM, VIO and Modern Map-Building in Shipping Hardware

Understanding SLAM & Localisation in Shipping Robotics

Simultaneous localisation and mapping (SLAM) is no longer a research novelty. It is a mature software layer that runs on commodity sensors and shipping compute modules across industrial AGVs, delivery platforms, and humanoid prototypes. The core problem remains unchanged: estimate robot pose relative to an environment while constructing a geometric or semantic map, using sensor fusion and optimisation. What has shifted is the hardware baseline. Modern SLAM stacks are evaluated by the sensors that ship, the compute that runs them, and the platforms that deploy them, not by renderings or press releases.

This article grades SLAM technologies by shipping hardware first, pilot deployments second, and announcements last. We cover feature-based localisation (ORB-SLAM), visual-inertial odometry (VIO), modern map-building pipelines, and the sensor ecosystems that power them. India availability and approximate INR landed costs are noted where components are commercially accessible.

Feature-Based Localisation: ORB-SLAM and Open-Source Foundations

ORB-SLAM remains the reference implementation for feature-based visual SLAM. The library tracks oriented FAST corners, computes BRIEF descriptors, and matches them across frames while maintaining a local map of keyframes and map points. Optimisation uses a sliding window bundle adjustment and loop closure detection to correct drift. The stack is framework-agnostic, runs on CPU or GPU, and is widely integrated into ROS/ROS2 navigation pipelines.

ORB-SLAM3 introduced multi-map support, sensor fusion (monocular, stereo, RGB-D, and IMU), and automatic initialisation. It does not require calibration to a fixed baseline and can switch between tracking modes at runtime. The code is open-source under GPL v3, which means commercial integration requires a separate license or fork management. Indian system integrators frequently use ORB-SLAM3 in ROS-based navigation stacks for warehouse inspection, facility mapping, and educational platforms.

ORB-SLAM3 Architecture and Real-World Deployment

Shipping hardware that runs ORB-SLAM3 typically pairs a depth or stereo camera with an ARM or x86 compute module. Examples include:

• Intel RealSense D435i or D455 (RGB-D + IMU) running on NVIDIA Jetson Orin Nano or Raspberry Pi Compute Module 4
• Trossen Robotics Pioneer 3DX or TurtleBot 4 with ORB-SLAM3 for indoor mapping and path planning
• Unitree Go2 and H1 platforms using custom ROS2 navigation nodes that incorporate feature tracking for texture-rich environments

ORB-SLAM3 performs well in structured indoor spaces with high texture contrast. It degrades in low-light, repetitive patterns, or highly reflective surfaces. For these cases, VIO or LiDAR fusion is required. The stack itself costs nothing to download, but integration support, commercial licensing, and hardware carry measurable costs.

Visual-Inertial Odometry (VIO) and Modern Map-Building

Visual-inertial odometry fuses camera frames with IMU measurements to estimate pose at high frequency. The IMU provides gravity alignment, motion prediction, and robustness to fast motion or brief occlusions. VIO pipelines typically use preintegration of IMU data, photometric or geometric error minimisation, and sliding-window optimisation. Popular open-source implementations include VINS-Fusion, OKVIS, and RTAB-Map with IMU support.

Modern map-building extends VIO beyond point clouds. Semantic segmentation, instance mapping, and submap management reduce storage and improve loop closure. Libraries like Cartographer, Fast-LIO2, and LIO-SAM handle LiDAR-heavy mapping, while VIO-centric stacks like VINS-Fusion and RTAB-Map handle camera-dominant pipelines. Multi-modal fusion (VIO + wheel odometry + LiDAR) is the shipping standard for humanoid and AGV platforms.

Hardware Platforms Running VIO Today

VIO requires a calibrated camera-IMU pair and a compute module capable of real-time optimisation. Shipping configurations include:

• NVIDIA Jetson Orin Nano (40 TOPS, 8 GB LPDDR5) running VINS-Fusion or RTAB-Map for indoor navigation
• Intel RealSense D435i (stereo depth + 6-axis IMU) paired with Jetson Orin or x86 SBCs
• STMicroelectronics ICM-42688-P or InvenSense ICM-20948 IMUs integrated into custom carrier boards for humanoid actuator joints
• Unitree H1 and G1 platforms using factory-tuned VIO+LiDAR fusion for dynamic locomotion and obstacle avoidance

Compute and sensor costs are the primary constraint. In India, Jetson Orin Nano modules are available through authorised distributors with landed costs around ₹28,000 to ₹32,000. Intel RealSense D435i units typically land between ₹16,000 and ₹19,000. IMUs like the ICM-42688-P cost ₹350 to ₹600 per unit when sourced via Digi-Key or Mouser. These estimates reflect GST, shipping, and distributor margins as of late 2024 and may vary by region and volume.

Multi-Modal Localisation and LiDAR Integration

LiDAR SLAM remains the baseline for outdoor and large-scale indoor mapping. 2D LiDAR stacks like Cartographer and Hector SLAM run on low-power compute, while 3D LiDAR pipelines use LOAM variants, FAST-LIO2, or LeGO-LOAM. Modern systems fuse LiDAR with VIO and wheel odometry to correct drift, handle dynamic objects, and maintain consistency across floors and staircases.

Shipping LiDAR options include Slamtec RPLIDAR A2/M2, Hesai Pandar16, and Ouster OS0/OS1 series. Ouster and Hesai provide SDKs that output calibrated point clouds with IMU timestamps, enabling direct integration into ROS2 navigation stacks. Indian integrators commonly deploy Slamtec RPLIDAR A2 for facility mapping and Ouster OS0-64 for outdoor or high-ceiling industrial environments.

LiDAR pricing in India reflects import duties and distributor margins. Slamtec RPLIDAR A2 units land around ₹9,000 to ₹11,000. Hesai Pandar16 and Ouster OS0-64 modules typically range from ₹1,10,000 to ₹1,60,000 landed, depending on calibration, mounting kits, and warranty terms. These costs are estimates and subject to currency fluctuation and customs policy.

India Availability, Component Pricing and Integration Notes

SLAM stacks are software. The hardware that runs them is widely available in India through authorised distributors, direct imports, and local robotics suppliers. The primary constraints are compute throughput, sensor calibration, and integration support.

Shipping Hardware and Pilot Deployments

• Boston Dynamics Spot: ships with factory-tuned VIO+LiDAR SLAM for inspection and mapping. Deployed in Indian industrial sites for asset monitoring.
• Unitree H1 and G1: humanoid platforms with factory-tuned VIO+LiDAR fusion. Pilots focus on dynamic locomotion, stair navigation, and obstacle avoidance.
• Agility Robotics Digit: shipping with multi-modal localisation for warehouse navigation. Pilots in logistics and manufacturing facilities.
• NVIDIA Jetson Orin-based custom platforms: widely used in Indian robotics labs and integrators for VIO, SLAM, and semantic mapping development.

Pilot deployments confirm that SLAM performance depends on environment texture, lighting, IMU quality, and compute scheduling. Factory-tuned stacks reduce integration risk. Open-source stacks require calibration, tuning, and continuous maintenance.

Commercial Licensing and Support in India

ORB-SLAM3, VINS-Fusion, and RTAB-Map are open-source. Commercial use requires license compliance or fork management. Indian system integrators typically secure support through:

• ROS2 navigation packages with SLAM contrib modules
• Local robotics consultancies offering calibration and tuning services
• Distributor technical support for Intel RealSense, Ouster, and Hesai SDKs

Integration timelines average 6 to 12 weeks for sensor calibration, VIO tuning, and loop closure validation. Hardware costs remain the dominant variable. Software licensing is secondary unless proprietary SLAM stacks are selected.

Conclusion

SLAM & localisation is a shipping-ready layer. ORB-SLAM3 provides a robust feature-based baseline for textured environments. VIO pipelines handle fast motion and low-texture conditions when paired with calibrated IMUs. Multi-modal fusion with LiDAR and wheel odometry is the shipping standard for humanoid and AGV platforms. India availability is strong for compute, cameras, and LiDAR, with landed costs clearly defined. The technology matures through deployment, not announcements. Hardware selection, calibration discipline, and integration support determine real-world performance.

References

• ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM. https://github.com/UZ-SLAMLab/ORB_SLAM3
• Intel RealSense D435i and D455 Datasheets and Developer Guides. https://www.intelrealsense.com/depth-camera-d435i/ and https://www.intelrealsense.com/depth-camera-d455/
• VINS-Fusion: A Robust and Versatile Monocular Visual-Inertial State Estimator. https://github.com/HKUST-Aerial-Robotics/VINS-Fusion
• NVIDIA Jetson Orin Nano Developer Kit Specifications. https://developer.nvidia.com/embedded/jetson-orin-nano
• Unitree Robotics H1 and G1 Platform Technical Briefs. https://www.unitree.com/products/h1 and https://www.unitree.com/products/g1
• Agility Robotics Digit Platform Deployment Information. https://agilityrobotics.com/digit
• Ouster OS0 LiDAR Datasheet and SDK Documentation. https://ouster.com/products/os0 and https://github.com/Ouster-SDK/ouster-sdk
• Slamtec RPLIDAR A2 Product Page and Specifications. https://www.slamtec.com/en/Lidar/A2
• Cartographer SLAM Library Documentation. https://google-cartographer.readthedocs.io/
• Fast-LIO2: Tightly-coupled GPU-accelerated LiDAR-inertial Odometry. https://github.com/gaoxiang12/Fast-LIO