Warehouse Automation Training Data

Claru collects warehouse automation data through both live-operation instrumentation and controlled test-station collection. For bin picking, we build SKU-diverse datasets spanning 500-5,000+ unique products with automated success/failure labeling from suction pressure and gripper force sensors. For palletizing, we capture full pallet-build sequences with box dimensions, placement coordinates, and stability assessments. Our warehouse collection protocol integrates with WMS systems through SKU identifier matching, enabling direct analysis of which product categories need additional training data. All episodes include cycle-time breakdowns by phase, grasp quality metrics, and failure taxonomy labels. We deliver in RLDS, HDF5, or custom formats with sensor calibration, SKU metadata linked to product databases, and stratified splits by SKU category, difficulty level, and warehouse zone. Typical turnaround is 4-8 weeks for datasets of 50,000+ pick episodes.

Warehouse automation encompasses the robotic systems that handle goods within distribution and fulfillment centers: goods-to-person retrieval (AMRs bringing shelves to pickers), autonomous pick-and-place from bins and shelves, automated packing (placing items into shipping boxes), palletizing (stacking boxes onto pallets in stable configurations), sorting (routing packages to outbound lanes), and last-mile delivery staging. The e-commerce logistics market processes billions of packages annually — Amazon alone ships over 5 billion packages per year in the US — and labor costs represent 65-70% of warehouse operating expenses. The global warehouse automation market exceeded $23 billion in 2024 and is projected to reach $41 billion by 2027.

Companies driving this space include Covariant (AI-powered bin picking), Berkshire Grey (intelligent robotic sorting), Dexterity (robotic palletizing and depalletizing), Locus Robotics and 6 River Systems (AMR-based goods-to-person), Nimble Robotics (robotic pick-and-pack), and RightHand Robotics (piece-picking for e-commerce). Amazon's Sparrow robot and Agility Robotics' Digit humanoid represent the vertically integrated approach. Despite billions in investment, most warehouse robots still handle only a fraction of SKUs (Stock Keeping Units) autonomously. The typical Amazon fulfillment center carries 100,000+ unique SKUs ranging from rigid boxes to deformable pouches to irregularly shaped items, and achieving 95%+ pick success across this diversity requires orders of magnitude more training data than current datasets provide.

The data challenge in warehouse automation is defined by SKU diversity and throughput pressure. A robot that achieves 99% pick success on 100 trained SKUs but fails on novel SKUs is commercially useless — warehouses cannot retrain the system every time a new product arrives. Generalization across SKU appearance (shape, size, weight, material, packaging) is the central technical challenge, and it requires training data that spans thousands of unique products. Simultaneously, warehouse robots must operate at commercial throughput rates (600-1,200 picks per hour for piece-picking, one pallet per 2-3 minutes for palletizing), meaning the training data must include not just successful picks but time-optimized motion trajectories that minimize cycle time while maintaining reliability.

Bin picking is the most data-intensive warehouse task and has seen the most learning-based progress. Covariant's AI platform uses a foundation model trained on millions of real pick attempts across customer sites, enabling zero-shot generalization to novel SKUs. Their approach — large-scale real-world data collection with automated labeling (success/failure from sensor readings, not human annotation) — demonstrates the data scaling law for warehouse manipulation: pick success rate increases log-linearly with training data volume, and crossing the commercially viable 99% threshold requires 100K+ labeled picks spanning thousands of SKUs. RightHand Robotics' RightPick system uses a suction-based gripper with learned grasp point selection, processing RGB-D point clouds to identify the best suction point on each item. Their published results show 95% first-attempt success on trained SKU categories, dropping to 85% on novel categories — a gap that closes only with additional training data.

Palletizing has traditionally used geometric planners that compute stable packing patterns from known box dimensions. However, real-world palletizing involves boxes that arrive in unpredictable order, are not always perfectly rectangular (damaged, overstuffed, or irregularly shaped), and must be stacked to maximize pallet stability for transport. Dexterity's AI palletizing system learns stacking policies from demonstrations that handle these real-world variations, using force/torque feedback to detect unstable placements and adjust. Their system requires 5,000-10,000 demonstrated palletizing sequences to learn robust stacking patterns across the range of box sizes and arrival orders encountered in a typical warehouse.

Autonomous mobile robots (AMRs) for goods-to-person represent a different data paradigm. Navigation data is primarily collected through fleet telemetry — every robot in a warehouse fleet generates hundreds of traversals per shift, providing abundant driving data. The challenge is not volume but annotation: extracting near-miss events, congestion patterns, and path efficiency metrics from raw telemetry. Locus Robotics and 6 River Systems operate fleets of thousands of AMRs in customer warehouses, and each robot continuously uploads navigation logs. Training data for AMR path planning comes from this fleet data, supplemented with simulation in digital twins of the warehouse floor plan using tools like NVIDIA Isaac Sim or Gazebo with warehouse-specific plugins.

The emerging frontier is multi-robot coordination. A modern fulfillment center may have 50-200 AMRs navigating simultaneously, plus 5-20 stationary picking and packing arms. Orchestrating this fleet to maximize throughput — avoiding congestion at popular aisles, balancing pick workload across stations, and handling robot failures gracefully — requires coordination data that captures fleet-level dynamics, not just individual robot behavior. This is the least explored area of warehouse automation data, and datasets that capture full-fleet telemetry with pick-station synchronization are extremely rare.

The sensor stack for warehouse picking combines vision, touch, and logistics sensors. For bin picking: overhead RGB-D cameras (Intel RealSense L515 or Photoneo PhoXi for high-resolution structured light) provide scene point clouds, a wrist-mounted camera gives close-up views of the grasp zone, suction cup pressure sensors or parallel-jaw force sensors indicate grip quality, and a barcode/RFID reader identifies the picked item for inventory tracking. For palletizing: overhead cameras capture the pallet loading pattern, a load cell under the pallet measures cumulative weight and center-of-mass shift, and box dimension sensors (typically a light curtain array) measure each box before stacking. For AMR navigation: 2D LiDAR (SICK TiM781 or Hokuyo URG-04LX) provides obstacle detection, wheel encoders give odometry, and optional 3D cameras provide vertical obstacle detection.

Collection in active warehouses is operationally complex. Most collection happens alongside live operations, meaning robots are picking real customer orders. This creates pressure to maintain throughput — the warehouse cannot slow down for data collection. The practical approach is instrumented operation: equip production robots with additional logging capability (higher-frequency sensor recording, video storage, extended telemetry) and capture data during normal operations. Alternatively, set up a dedicated test station with a representative selection of SKUs for controlled data collection during off-peak hours. Claru supports both modes: live-operation instrumentation for high-volume, diverse data, and controlled-station collection for systematic coverage of edge cases (fragile items, oddly shaped products, multi-item picks).

SKU diversity is the most critical data axis and the most expensive to achieve. A production bin-picking dataset must include: rigid boxes (varying sizes from envelope to shoebox), poly-bagged items (deformable, reflective, hard to suction), bottles and cylinders (roll-prone), small items (pills, screws — hard to detect and grasp), heavy items (books, tools — require force-limited motion), and multi-packs (shrink-wrapped bundles). Each SKU category requires 50-200 pick attempts to adequately cover the range of bin positions, orientations, and clutter states. For a 1,000-SKU dataset with 100 attempts per SKU, this is 100,000 pick recordings — a major collection effort that justifies professional data services.

Annotations for warehouse data follow the logistics workflow. Each pick episode is labeled with: SKU identifier (matched to warehouse inventory system), pick success/failure with failure taxonomy (suction seal failure, weight exceeded, item dropped, wrong item picked, bin collision), cycle time (approach to post-place), grasp quality metrics (suction vacuum level or jaw force at lift), and item condition (damaged during pick, label obscured). For palletizing, each box placement is annotated with: box dimensions and weight, placement coordinates on the pallet, stacking stability prediction, and pattern compliance. All annotations are linked to the warehouse management system (WMS) through order ID and timestamp, enabling offline analysis of which SKU categories have the highest failure rates and need additional training data.