SLAM & Localisation: The Architecture Behind Autonomous Navigation

SLAM & Localisation: The Architecture Behind Autonomous Navigation

Simultaneous Localisation and Mapping (SLAM) remains the foundational perception stack for robots operating in GPS-denied or unstructured environments. For humanoid platforms, mobile manipulators, and warehouse automation hardware, localisation accuracy directly dictates task reliability. This article grades claims by shipping hardware first, pilot deployments second, and announcements last. We focus on ORB-SLAM, visual-inertial odometry (VIO), and modern map-building pipelines, separating academic benchmarks from deployed reality.

From EKF Filters to Optimization-Based Tracking

Early SLAM systems relied on Extended Kalman Filters (EKF-SLAM), which linearized non-linear sensor models and propagated uncertainty through covariance matrices. While computationally light, EKF approaches struggled with scale ambiguity, loop closure errors, and large-scale environments. The shift to optimization-based SLAM introduced factor graphs and bundle adjustment, allowing robots to jointly optimize camera poses, IMU states, and map points in a single nonlinear least-squares problem.

Modern optimization pipelines use sliding-window or keyframe-based formulations. Instead of maintaining a dense covariance matrix over all historical states, these systems marginalize older variables while preserving geometric constraints. This reduces memory footprint and improves real-time performance on embedded compute. The trade-off is increased implementation complexity and sensitivity to initialization, outlier rejection, and IMU calibration drift.

ORB-SLAM: Open-Source Frameworks in Production Contexts

ORB-SLAM2 and ORB-SLAM3 emerged as reference implementations for feature-based monocular, stereo, and RGB-D SLAM. The pipeline divides into three concurrent threads: tracking (visual odometry with IMU fusion in VIO mode), local mapping (keyframe insertion, point cloud optimization, and covisibility graphs), and loop closing (word-tree recognition and pose graph optimization).

Grading claims by hardware deployment: ORB-SLAM is widely cited in research and open-source robotics, but production systems rarely ship with unmodified ORB-SLAM. Manufacturing teams typically fork the framework, replace ORB features with more robust descriptors (e.g., AKAZE, SuperPoint), integrate custom IMU preintegration, and harden outlier rejection for dynamic scenes. The academic benchmarks on EuRoC and TUM datasets show sub-meter accuracy under controlled conditions, but real-world deployment requires calibrated extrinsics, temperature-compensated IMUs, and texture-rich or structured lighting environments.

Shipping hardware that implements ORB-SLAM-derived stacks includes custom-built mobile manipulators and research-grade humanoid prototypes. OEMs often license modified versions or build proprietary forks to meet reliability, support, and licensing requirements. Announcements claiming "ORB-SLAM powered" should be cross-referenced with factory videos, SDK documentation, and pilot deployment reports before treating them as production-ready.

Visual-Inertial Odometry: Sensor Fusion for GPS-Denied Environments

VIO combines frame-to-frame camera motion with IMU preintegration to resolve scale, reduce drift, and improve temporal alignment. The IMU provides high-frequency linear acceleration and angular velocity, while the camera supplies metric scale and loop closure. For humanoids, VIO is critical during legged locomotion where foot-ground contact introduces high-frequency vibration and motion blur.

Production VIO stacks typically use: (1) a rolling-shutter or global-shutter camera with known focal length and distortion coefficients, (2) a MEMS IMU calibrated for bias, scale factor, and cross-axis misalignment, (3) a sliding-window optimizer (e.g., VINS-Fusion, OKVIS, or custom implementations), and (4) a pose graph backend for loop closure. The pipeline must handle: (a) motion blur during rapid arm swings, (b) lighting changes across indoor-outdoor transitions, (c) textureless surfaces like white walls or glass, and (d) dynamic objects that violate static-world assumptions.

Grading by deployment tier: Shipping hardware from established robotics suppliers includes integrated VIO modules with factory calibration certificates. Pilot deployments in Indian logistics and manufacturing facilities commonly use stereo RGB-D cameras paired with Jetson Orin Nano or Thor compute, running tuned VIO stacks. Announcements of "VPS-ready" or "autonomous navigation" claims should be verified against independent demo videos, sensor spec sheets, and third-party integration reports.

Modern Map-Building: Sparse, Dense, and Semantic Representations

Map representations determine how robots store, retrieve, and update spatial information. Sparse maps store keyframes and feature points, offering low memory usage and fast loop closure. Dense maps reconstruct voxel or mesh surfaces, enabling collision avoidance and fine-grained manipulation planning. Semantic maps attach class labels to geometric structures, supporting language-guided navigation and task planning.

Production systems prioritize sparse or semi-dense maps for navigation, reserving dense reconstruction for inspection or manipulation tasks. Neural implicit representations (NeRF, 3D Gaussian Splatting) show promise in reconstruction fidelity, but shipping humanoid hardware rarely uses them for real-time localisation due to inference latency, memory constraints, and lack of standardized pose-graph integration. These techniques remain in pilot or research stages, with deployment timelines dependent on edge-optimized renderers and standardized SLAM backends.

Map-building pipelines in deployed hardware typically include: (1) feature extraction and matching, (2) depth estimation via stereo or ToF, (3) point cloud registration (ICP/NDT), (4) pose graph optimization with loop closure, and (5) map compression for long-term storage. Teams must balance accuracy, update rate, and memory budget. Commercial VPS (Visual Positioning System) providers offer cloud-assisted localisation, but latency and subscription costs limit use in safety-critical or offline environments.

India Availability, Pricing, and Deployment Reality

India's robotics supply chain for SLAM and localisation has matured significantly, though component sourcing still relies on imports for high-end sensors and compute modules. Approximate landed costs (INR, as of 2024) include:

• Intel RealSense D455/D435i (RGB-D + IMU): ₹18,000–₹25,000 per unit
• Luxonis OAK-D-Lite / OAK-D Pro (stereo + AI accelerator): ₹12,000–₹35,000 per unit
• STMicroelectronics ICM-42688-P / Bosch BNO086 IMU: ₹800–₹2,500 per unit
• NVIDIA Jetson Orin Nano / Orin NX (compute modules): ₹45,000–₹95,000 per module
• Custom VIO calibration rigs (factory or third-party): ₹15,000–₹40,000 per setup

SLAM software stacks are predominantly open-source (ORB-SLAM3, VINS-Fusion, RTAB-Map, Cartographer). Commercial licensing varies by vendor, with OEM packages typically ranging from ₹50,000 to ₹3,00,000 per unit for enterprise support, calibration tools, and long-term maintenance. Visual Positioning System (VPS) subscriptions in India average ₹2,000–₹8,000 per robot per month, depending on coverage density and API tier.

Pilot deployments in Indian warehouses, hospitals, and manufacturing facilities show that localisation accuracy degrades under: (1) low-texture corridors, (2) dynamic crowd movement, (3) thermal drift in unconditioned environments, and (4) improper IMU-camera time synchronization. Shipping hardware with factory calibration, hardware timestamping, and redundant sensors (e.g., LiDAR + VIO) consistently outperforms software-only announcements. Teams should prioritize integration testing over feature lists when evaluating localisation stacks.

References

1. ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM. GitHub Repository. https://github.com/UZ-SLAMLab/ORB_SLAM3

2. Intel RealSense D400 Series Depth Cameras Datasheet. Intel Corporation. https://www.intelrealsense.com/depth-cameras/d435i/

3. Luxonis OAK-D Series Product Specifications. Luxonis Inc. https://docs.luxonis.com/software/samples/depthai-examples/

4. VINS-Fusion: A Robust and Versatile Monocular Visual-Inertial State Estimator. IEEE Transactions on Robotics, 2018. https://arxiv.org/abs/1708.03852

5. EuRoC MAV Dataset. ETH Zurich. https://projects.asl.ethz.ch/datasets/doku.php?id=kmavvisualinertialdatasets

6. NVIDIA Jetson Orin Series Developer Specifications. NVIDIA Corporation. https://developer.nvidia.com/embedded/jetson-orin

7. RTAB-Map: Real-Time Appearance-Based Mapping. Université Laval. https://github.com/introlab/rtabmap

8. Cartographer: Mapping in Real-Time. Google Research. https://github.com/cartographer-project/cartographer