Navigation Training Data

Claru collects navigation data using standardized mobile platforms deployed across our global network of 100+ cities, providing the environment diversity that is the primary driver of navigation policy generalization. Each recording includes hardware-synchronized RGB, LiDAR, IMU, and odometry streams with ground-truth poses from visual-inertial odometry or RTK-GPS. Our collection protocol captures diverse indoor and outdoor environments with natural human traffic at multiple times of day, ensuring coverage of lighting variation, dynamic obstacle densities, and terrain types. We deliver processed datasets with semantic map annotations, traversability labels, dynamic obstacle bounding boxes with velocity vectors, and human-verified trajectory quality scores. Standard delivery includes 50-200 hours of navigation data per deployment domain, formatted for direct ingestion by ViNT, GNM, NoMaD, or custom navigation architectures in RLDS, HDF5, or ROS bag formats.

Robot navigation — the ability to move autonomously from one location to another while avoiding obstacles — is the foundational capability that enables every downstream mobile robotics application. Whether a robot is delivering packages in a warehouse, assisting patients in a hospital, or patrolling a construction site, it must first solve navigation reliably. The mobile robot market is projected to exceed $54 billion by 2030, and while classical navigation stacks (SLAM + A* + DWA) work in controlled environments, they fail systematically in the unstructured, dynamic real world where learned navigation policies are required.

The fundamental challenge is perceptual diversity. A navigation policy must handle reflective floors that confuse LiDAR, transparent glass walls invisible to depth sensors, dynamic obstacles like pedestrians who change direction unpredictably, and lighting conditions ranging from direct sunlight to pitch darkness. Classical cost maps cannot encode these nuances — they treat all obstacles identically and all free space as equally traversable. Learned policies trained on diverse real-world trajectory data implicitly capture the full complexity of navigable environments, including soft preferences like staying away from fragile objects or yielding to pedestrians.

The Visual Navigation Transformer (ViNT, Shah et al., 2023) demonstrated that a single navigation policy trained on diverse trajectory data from 6 different robot platforms totaling 100+ hours of experience can generalize to entirely new environments zero-shot. The key insight was data diversity over data volume: ViNT's cross-embodiment training set spanning indoor hallways, outdoor sidewalks, and off-road trails produced stronger generalization than any single-environment dataset regardless of size. Similarly, GNM (Shah et al., 2023) showed that goal-conditioned navigation improves with environment diversity more than raw trajectory count.

Policies trained exclusively in simulation show 20-40% performance degradation in real deployment due to the sim-to-real gap in perceptual inputs. Simulated environments cannot capture real floor textures (reflections, scuff marks, wet patches), realistic sensor noise patterns (LiDAR multipath in narrow corridors, depth sensor failures on dark surfaces), or the behavioral patterns of real dynamic obstacles. For production deployment, real-world navigation data collected across the target environment distribution is not optional — it is the primary determinant of deployment reliability.

GNM (Shah et al., 2023) introduced the general navigation model paradigm: a single goal-conditioned policy trained on trajectories from multiple robots that transfers zero-shot to new platforms. Trained on 2,735 trajectories across 6 environments from 3 different robot platforms, GNM achieved 85% goal-reaching success on unseen environments versus 42% for a policy trained on a single environment. The architecture processes a current observation and goal image to predict a waypoint sequence, decoupling perception from low-level control and enabling cross-embodiment transfer.

ViNT (Shah et al., 2023) scaled this approach with a transformer backbone trained on over 100 hours of navigation data from 6 robot embodiments spanning indoor, outdoor, and off-road domains. ViNT demonstrated that navigation is a scalable learning problem: performance on held-out environments improved log-linearly with training data diversity, reaching sub-30cm goal accuracy on environments never seen during training. The model uses an EfficientNet visual encoder with a GPT-style action decoder that predicts 8-step future waypoints at 4 Hz.

NoMaD (Sridhar et al., 2023) extended foundation navigation models with diffusion-based action prediction, achieving more robust behavior in cluttered environments where multimodal action distributions matter. NoMaD's diffusion head predicts a distribution over future trajectories rather than a single path, naturally handling decision points like choosing which side of an obstacle to pass. On the RECON benchmark, NoMaD reduced collision rate by 43% compared to ViNT while maintaining comparable goal-reaching performance.

LM-Nav (Shah et al., 2023) demonstrated that large language models can provide the high-level planning layer for navigation without any navigation-specific language training. By combining a pre-trained visual navigation model with GPT-4 for instruction parsing and CLIP for landmark grounding, LM-Nav executed complex multi-step navigation instructions (e.g., 'go past the fountain, turn left at the red building, and stop at the bench') with 82% success rate in outdoor campus environments — entirely from pre-trained models without navigation-specific fine-tuning.

Production navigation data collection requires instrumented mobile platforms that capture synchronized multi-sensor streams while traversing target environments. The core sensor suite includes a front-facing RGB camera (minimum 640x480 at 30 Hz), a 2D or 3D LiDAR scanner (10-20 Hz), wheel odometry or visual-inertial odometry for ego-motion estimation, and an IMU (100-200 Hz) for dead-reckoning through GPS-denied areas. For outdoor navigation, RTK-GPS provides centimeter-level ground truth positioning. All sensors must be hardware-synchronized and extrinsically calibrated to a common body frame.

Environment diversity is the single most important variable for navigation dataset quality. A production dataset should span at least 20 distinct environments covering the target deployment domain. For indoor service robots, this means offices with open floor plans and cubicle mazes, hospitals with long corridors and elevator lobbies, retail stores with dense aisle layouts, and homes with narrow doorways and furniture clutter. Each environment needs trajectories from 50+ distinct start-goal pairs to ensure spatial coverage, with recordings at different times of day to capture lighting variation.

Dynamic obstacle coverage is critical and often underrepresented in navigation datasets. At least 30% of trajectories should include active obstacle avoidance events — pedestrians crossing the path, doors opening, carts being moved. Record these as they occur naturally rather than staging encounters, because the distribution of human movement patterns (speed, predictability, density) varies dramatically between hospitals, offices, and retail environments. Each trajectory should be annotated with obstacle encounter timestamps and avoidance outcomes.

Claru collects navigation data using standardized mobile platforms deployed across our global network of 100+ cities. Each recording includes synchronized RGB, LiDAR, IMU, and odometry streams with ground-truth poses from visual-inertial odometry or RTK-GPS. We capture diverse indoor and outdoor environments with natural human traffic, delivering processed datasets with semantic map annotations, traversability labels, dynamic obstacle bounding boxes with velocity vectors, and human-verified trajectory quality scores. Our standard delivery includes 50-200 hours of navigation data per deployment domain.