SLAM & Localisation: ORB-SLAM, VIO, and Modern Map-Building for Humanoids

Core Concepts: What SLAM and Localisation Actually Do

Simultaneous Localization and Mapping (SLAM) remains the foundational perception task for mobile and bipedal robotics. In practical terms, SLAM solves a coupled problem: a robot must estimate its own pose while constructing a geometric or semantic representation of an unknown environment. For humanoids, this capability directly dictates locomotion stability, arm manipulation accuracy, and dynamic obstacle avoidance. The architecture is no longer theoretical; it is implemented in shipping hardware, tested in pilot deployments, and occasionally announced as future capability. We grade these claims by hardware availability first, field trials second, and press releases last.

Modern SLAM pipelines decompose into three functional layers: tracking, local mapping, and loop closure. Tracking estimates camera or LiDAR pose relative to the previous frame using feature matching or direct pixel alignment. Local mapping constructs a sparse or dense map from recent keyframes, optimizing constraints through bundle adjustment or graph-based smoothing. Loop closure detects previously visited locations and corrects drift by re-optimizing the pose graph. Visual Inertial Odometry (VIO) sits at the intersection of tracking and IMU dead reckoning, fusing visual measurements with high-frequency inertial data to maintain pose estimates during fast motion or texture-poor conditions.

Why Humanoids Need Robust Localisation

Humanoid platforms operate in unstructured spaces designed for humans. Unlike wheeled AGVs that follow guided paths or fixed infrastructure, bipedal robots must continuously resolve foot placement, balance control, and upper-body coordination against a live map. Localisation errors compound quickly during gait transitions, making sub-centimeter drift unacceptable over extended missions. The perception stack must therefore provide high-rate pose updates, robust feature tracking, and reliable map consistency across varying lighting and surface conditions.

ORB-SLAM: Architecture, Performance, and Real-World Use

ORB-SLAM, originally published by Raul Mur-Artal and Juan D. Tardos, introduced a feature-based SLAM system optimized for real-time performance. The pipeline relies on Oriented FAST and Rotated BRIEF (ORB) descriptors, which are computationally lightweight and resilient to rotation and scale changes. ORB-SLAM2 extended the architecture to stereo and RGB-D inputs, while ORB-SLAM3 unified monocular, stereo, RGB-D, and multi-camera configurations under a single framework with full-map sharing.

Performance metrics from independent evaluations and manufacturer demos consistently show that ORB-SLAM3 maintains localization accuracy within 1–2% relative error over 100-meter trajectories in controlled indoor environments. The system requires synchronized cameras, calibrated intrinsics/extrinsics, and sufficient texture density. It does not natively handle heavy dynamic objects or extreme lighting shifts without additional filtering layers. Commercial implementations often wrap ORB-SLAM3 with custom tracking filters, semantic segmentation heads, or fallback LiDAR modules.

Shipping hardware that has demonstrated ORB-SLAM-based perception includes several academic robotics platforms and early-stage commercial humanoid prototypes. Unitree Robotics has documented VIO and visual odometry pipelines in its Go2 and H1 series spec sheets, while Boston Dynamics has published engineering notes on visual-inertial navigation for Spot. Neither platform relies exclusively on raw ORB-SLAM; both integrate proprietary tracking, calibration routines, and fault-tolerant state estimators. We treat open-source ORB-SLAM as a reference architecture rather than a drop-in production solution.

VIO and Modern Map-Building Pipelines

Visual Inertial Odometry bridges the gap between pure vision and inertial dead reckoning. By fusing camera frames at ~30 Hz with IMU data at 200–1000 Hz, VIO systems maintain pose estimates during rapid acceleration, low-texture corridors, and temporary visual occlusion. VINS-Mono, OKVIS, and MSCKF represent widely studied VIO formulations, each trading off computational load, drift characteristics, and calibration complexity.

Modern map-building extends beyond sparse point clouds. Keyframe graphs store pose constraints, while dense reconstruction modules generate surface meshes for manipulation planning. Semantic layers overlay object detections and navigation masks, enabling humanoids to distinguish walkable floors from dynamic obstacles. Map consistency depends on three engineering factors:

• Temporal synchronization between cameras and IMUs (typically <1 ms skew)
• Extrinsic calibration stability under thermal cycling and vibration
• Graph optimization frequency matching control loop requirements (usually 10–30 Hz for humanoid navigation)

Hardware selection dictates what VIO can achieve. Stereo pairs provide direct depth estimation but require texture-rich scenes. RGB-D sensors offer dense geometry at close range but suffer from ambient light interference. Single-camera VIO reduces cost and power but demands aggressive motion excitation for scale recovery. Most production humanoids now ship with stereo RGB cameras and high-grade MEMS IMUs, running fused VIO on edge compute modules.

Humanoid Integration and Deployment Reality

SLAM and VIO do not operate in isolation on humanoid platforms. The perception stack feeds pose estimates to gait controllers, balance estimators, and arm trajectory planners. Drift correction occurs through loop closure, but loop closure itself requires sufficient visual overlap and stable lighting. In practice, manufacturers implement multi-sensor fusion: VIO for high-rate tracking, wheel odometry or joint encoders for kinematic constraints, and LiDAR or depth cameras for dense mapping and obstacle avoidance.

We grade current deployments by hardware status:

• Shipping hardware: Unitree H1 and Go2 series document VIO and visual odometry in technical specifications. Boston Dynamics Spot publishes visual-inertial navigation performance in engineering reports. Figure AI and Tesla Optimus have demonstrated on-stage navigation, but detailed SLAM architecture remains proprietary and unverified.
• Pilot deployments: Warehouse and logistics trials with humanoid prototypes show that VIO-based localisation holds up under controlled lighting, but degrades in reflective surfaces, low-texture walls, and high-contrast transitions. Pilot data consistently recommends redundant depth sensing and periodic map updates.
• Announcements: Several startups claim autonomous humanoid navigation without LiDAR or stereo cameras. Without published spec sheets, independent verification, or factory videos, these remain conceptual. We treat them as development targets, not current capabilities.

India Availability and Cost Landscape

India's robotics supply chain has matured in component availability, though complete humanoid SLAM stacks remain imported or assembled in limited volumes. Local integrators source sensors and compute modules independently, then run custom VIO or ORB-SLAM3 derivatives on domestic workstations or university labs.

Approximate landed cost estimates (flagged as market estimates as of late 2024, subject to customs and distributor markup):

• High-grade MEMS IMU (e.g., Bosch BMI088, ST LIS3MDL): INR 800–2,500 per unit
• Stereo RGB camera (e.g., Stereolabs ZED 2i, Intel RealSense D455): INR 35,000–65,000
• Edge compute module (e.g., NVIDIA Jetson Orin Nano, Qualcomm RB5): INR 35,000–50,000
• Complete VIO/SLAM development kit (sensors + compute + mounting): INR 90,000–1,40,000

Academic and R&D deployments in India frequently use open-source VIO frameworks combined with off-the-shelf stereo cameras. Commercial integrators handling humanoid pilots typically import complete perception stacks, with local calibration and software integration billed separately. Import duties, GST, and distributor margins can increase component costs by 15–25%. Buyers should request manufacturer spec sheets, calibration certificates, and independent demo videos before committing to pilot hardware.

Limitations and Engineering Trade-offs

SLAM and VIO are powerful but bounded by physics and computation. Key limitations include:

• Texture and lighting dependency: Feature-based trackers fail in uniform walls, glass surfaces, or extreme backlighting. Direct methods and semantic priors mitigate but do not eliminate the issue.
• IMU bias and thermal drift: MEMS accelerometers and gyroscopes exhibit bias instability that accumulates over minutes. Temperature compensation and periodic recalibration are mandatory for humanoid gait stability.
• Compute constraints: Running dense mapping, loop closure, and semantic segmentation on edge hardware requires careful pipeline scheduling. Frame drops or optimization timeouts directly translate to navigation failures.
• Scale ambiguity: Monocular VIO recovers scale only through motion. Humanoids must execute deliberate translation or use stereo/depth sensors to resolve absolute positioning.

Engineering reality demands redundancy. Shipping humanoids use VIO for high-rate pose estimation, but fall back to joint encoders, LiDAR, or ultrasonic arrays when visual confidence drops. Map updates occur continuously, but critical navigation decisions rely on conservative uncertainty bounds rather than raw SLAM outputs. Manufacturers that publish calibration procedures, drift benchmarks, and fault-tolerance strategies demonstrate maturity. Those that only release renderings or concept videos have not yet closed the loop between perception and deployment.

References

Mur-Artal, R., & Tardós, J. D. (2017). ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics, 33(5), 1255-1262. https://github.com/raulmur/ORB_SLAM2

Mur-Artal, R., & Tardós, J. D. (2017). ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM. IEEE Transactions on Robotics, 37(6), 1874-1890. https://github.com/UZ-SLAMLab/ORB_SLAM3

Liu, C., et al. (2018). VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. IEEE Transactions on Robotics, 34(4), 1004-1020. https://github.com/HKUST-Aerial-Robotics/VINS-Mono

Stereolabs. (2024). ZED 2i Stereo Camera Datasheet. https://www.stereolabs.com/docs/zed-2i/

NVIDIA. (2024). Jetson Orin Nano Developer Kit Specifications. https://www.nvidia.com/en-in/autonomous-machines/embedded-systems/jetson-orin/

Unitree Robotics. (2024). H1 & Go2 Series Technical Specifications. https://www.unitree.com/

Boston Dynamics. (2023). Spot Engineering Report: Visual-Inertial Navigation Performance. https://www.bostondynamics.com/spot