Packing Task Training Data

Claru operates packing data collection stations that replicate real fulfillment center workflows with high-fidelity multi-sensor instrumentation. Each station features a Zivid structured-light sensor for sub-millimeter container heightmaps, a synchronized multi-camera array (2 overhead + 2 side RGB-D at 30 Hz), and packing material dispensers for realistic void fill demonstrations. We collect demonstrations on real product inventory supplied by clients, covering rigid boxes, deformable clothing, fragile electronics, and irregular consumer goods. Operators follow standardized packing protocols with quality scoring on volume utilization (75%+ target), stability, and damage prevention. Deliverables include per-placement heightmaps, 6-DoF item poses, placement sequences, item dimension annotations, and volume utilization scores — formatted for direct ingestion by 3D bin packing RL environments, Diffusion Policy, or custom sequence planning architectures. Daily throughput of 200-500 complete order demonstrations enables rapid dataset scaling for production packing systems.

Robotic packing — arranging items into containers for shipping or storage — is the mirror image of bin picking and arguably the harder problem. While picking requires selecting one item from clutter, packing requires placing multiple items into a constrained volume in a stable, damage-free configuration that minimizes void space. Amazon ships over 5 billion packages annually, and the packing station remains one of the most labor-intensive nodes in fulfillment operations. The 3D bin packing problem is NP-hard even for rigid cuboids; real-world packing with irregular shapes, fragile items, and deformable packaging materials like bubble wrap and air pillows is orders of magnitude more complex.

The data challenge in packing is fundamentally about spatial reasoning under constraints. A packing policy must simultaneously reason about item placement order (heavy items first, fragile items on top), geometric fit (maximizing volume utilization while maintaining stability), packing material insertion (void fill, separators, cushioning), and box flap closure or tape sealing. Human packers solve these constraints through years of spatial intuition — they can look at a set of items and mentally simulate stable arrangements in seconds. Transferring this spatial reasoning to robots requires demonstrations that capture not just the final placement poses but the entire decision process: which item to place next, where to position it, how to adjust neighboring items, and when to add protective packaging.

Current robotic packing systems rely heavily on heuristic algorithms that work for uniform products but fail on the mixed-item orders that dominate e-commerce. The Online 3D Bin Packing Problem benchmark (Zhao et al., 2022) showed that reinforcement learning policies trained on 500K+ episodes can achieve 72.1% volume utilization on random cuboid sets — competitive with the best-first-fit-decreasing heuristic at 68.5% — but real warehouse items are not cuboids. Deformable items like clothing, bags, and pouches cannot be modeled as rigid bodies, and their placement behavior depends on how they drape and compress against neighboring items. Learning these physics from demonstration data is currently the only viable path to general packing policies.

The economic incentive is enormous. McKinsey estimates that warehouse labor costs account for 65% of total fulfillment operating expenses, and packing stations require 2-3 workers per shift in a typical facility. Dimensional weight pricing by carriers means that poor packing directly increases shipping costs — a 10% improvement in volume utilization can save $0.15-$0.50 per package at scale. For a fulfillment center shipping 100,000 packages daily, that translates to $5.5-$18.3 million in annual savings from packing optimization alone, before accounting for labor cost reduction.

The Online 3D Bin Packing problem has been the primary benchmark for learned packing policies. Zhao et al. (2022) formulated the problem as a Markov Decision Process where items arrive one at a time and the agent must place each item before seeing the next. Their PCT (Packing Configuration Transformer) architecture achieved 72.1% volume utilization on sequences of 50 random cuboids, outperforming all heuristic baselines. The key insight was representing the container state as a heightmap and using attention over candidate placement positions, allowing the policy to reason about global packing quality rather than greedy local placement.

For real-world packing with irregular items, simulation alone is insufficient. Ha et al. (2024) demonstrated a sim-to-real pipeline for packing grocery items into bags, training in IsaacGym with 3D-scanned item meshes and transferring to a Franka Panda arm. The system achieved 82% packing success on 20 common grocery items, but performance dropped to 61% on deformable items (bread bags, chip bags) where simulation contact models diverge significantly from reality. This 21-percentage-point gap on deformable items underscores the need for real-world demonstration data that captures the actual physics of soft item packing.

PackIt (Goyal et al., 2020) explored learning geometric packing directly from 3D shape understanding. Given a set of items and a container, PackIt predicts a packing configuration by iteratively selecting items and placement poses using a learned shape-conditioned policy. On a benchmark of 12 household object categories, PackIt achieved 64% valid packing configurations versus 41% for a random placement baseline. However, PackIt operates on known 3D models and does not handle the perception challenge of estimating item geometry from sensor data in a cluttered staging area.

The most promising recent direction combines large vision-language models with physical reasoning. SpatialVLM (Chen et al., 2024) demonstrated that fine-tuning a VLM on spatial relationship annotations enables 3D spatial reasoning from 2D images — predicting relative positions, containment relationships, and stability assessments. While not yet applied specifically to packing, the spatial reasoning capabilities (83% accuracy on relative position questions, 76% on stability prediction) suggest that VLM-based packing planners could leverage pretrained spatial understanding with task-specific fine-tuning on packing demonstrations.

Packing data collection requires a staged workspace that mirrors real fulfillment station ergonomics. The setup includes an item staging area (conveyor or tote) where products arrive, a box selection station with multiple container sizes, packing material dispensers (air pillows, kraft paper, bubble wrap), and an overhead + side camera array to capture the evolving container state from multiple viewpoints. Depth sensing is critical for tracking the evolving heightmap inside the container — structured-light sensors (Zivid, Photoneo) mounted 60-80 cm above the box opening provide 0.2 mm accuracy across the packing volume.

Each packing demonstration captures the full order fulfillment sequence: box selection based on item set, item retrieval from the staging area, placement into the container with chosen orientation and position, void fill insertion, and box closure. Annotations must include item identity and measured dimensions, placement pose (6-DoF relative to the container), placement order within the sequence, contact state with neighboring items, void fill type and placement, and container fill level (percentage of volume utilized). For deformable items, additional annotations capture the deformation state — compressed height versus free-standing height, draping contact area, and whether the item was folded or compressed during placement.

Operator training for packing data collection focuses on consistent high-quality packing that balances volume utilization with item protection. We define packing quality metrics: volume utilization (target 75%+ for mixed items), stability (no item shifts when the box is tilted 15 degrees), damage prevention (fragile items cushioned, heavy items on bottom), and presentation (items visible and identifiable for quality inspection). Operators follow a standardized protocol: assess items, select box, place heaviest items first on the bottom layer, fill gaps with medium items, add cushioning around fragile items, place lightweight items on top, and add void fill. Each operator completes a 50-order qualification round with quality scoring before production collection begins.

For maximum diversity, packing sessions rotate through multiple order profiles: single-item orders (baseline placement data), multi-item homogeneous orders (stacking and layering patterns), multi-item heterogeneous orders (mixed-category spatial reasoning), orders with fragile items (protective packing strategies), and orders with deformable items (clothing, pouches). Each order profile contributes distinct aspects of packing intelligence. Single-item data teaches optimal item orientation and box selection; multi-item data teaches spatial sequencing and stability reasoning; fragile-item data teaches protective placement heuristics.

Claru operates packing data collection stations designed to replicate real fulfillment center workflows with instrumentation for high-fidelity demonstration capture. Each station features a multi-camera array (2 overhead + 2 side-mounted RGB-D sensors synchronized at 30 Hz) covering the full packing workspace from box selection through closure. A Zivid structured-light sensor mounted directly above the container captures sub-millimeter heightmaps of the evolving packing state after each item placement, providing the ground-truth spatial data needed for heightmap-based policies.

We collect packing demonstrations on real product inventory supplied by clients, covering the full spectrum of item categories: rigid boxes, cylindrical containers, flexible pouches, clothing, fragile electronics, and irregularly shaped consumer goods. Our operators are trained on standardized packing protocols with quality scoring on volume utilization, stability, and damage prevention. Each demonstration captures the complete sequence from box selection through item placement, void fill insertion, and closure — annotated with per-item placement poses, contact states, packing material usage, and fill level progression.

Claru delivers packing datasets formatted for direct ingestion by 3D bin packing RL environments, Diffusion Policy architectures, or custom sequence planning models. Standard deliverables include per-placement heightmaps, 6-DoF item poses relative to the container, placement order sequences, item dimension annotations, void fill placement data, and final volume utilization scores. For clients building end-to-end packing systems, we provide full manipulation trajectories with proprioceptive data covering the grasp-transport-place cycle for each item. Our daily throughput of 200-500 complete order packing demonstrations enables rapid dataset scaling for production packing policy training.