OptiPack

THE PROBLEM: SHIPPING AIR & THE $15,000 HARDWARE BARRIER IN E-COMMERCE LOGISTICS Every day, millions of multi-item e-commerce orders are packed by human guesswork. Warehouse operators visually estimate box sizes for heterogeneous items, resulting in two massive inefficiencies: 1. Dimensional (DIM) Weight Penalties & "Shipping Air": Average e-commerce cartons contain 40–60% empty void space. Carriers bill by volumetric weight (L×W×H / 5000), inflating shipping costs by 20–35% and wasting tons of corrugated cardboard and plastic dunnage. 2. Transit Damage from Improper Stacking: Fragile items (glassware, electronics, cosmetics) are routinely crushed under heavier parcels or rotated onto unstable axes because packers lack systematic 3D load-bearing guidance. While enterprise hubs use laser/ToF dimensioning gates (e.g., CubiScan), these systems cost $15,000–$50,000 per station and only measure one static item at a time—making them inaccessible to 99% of SMB merchants, dark stores, and decentralized 3PL fulfillment centers. Conversely, naive "AI wrapper" apps fail in production because Vision LLMs hallucinate metric dimensions from uncalibrated photos and cannot guarantee collision-free 3D physical geometry. THE SOLUTION: OPTIPACK OptiPack is an end-to-end 3D Metrology, Semantic Physics, and Constrained Cartonization Copilot that transforms any standard smartphone or webcam—paired with a free printable A4 ArUco sheet—into a sub-centimeter multi-item 3D dimensioner and optimal packing planner. From just two photos (one top-down zenith capture and one 35°–45° oblique capture), OptiPack executes a hybrid neuro-symbolic pipeline in under 2 seconds: 1. Calibrated Two-View 3D Metrology: Solves exact 6-DOF camera extrinsics from the 4-corner A4 ArUco reference board, segments all items autonomously using Meta SAM 3.1 (sam-3.1), rejects 2D floor shadows/reflections via Planar Parallax verification, and reconstructs true metric 3D bounding cuboids (L × W × H in mm) via Visual Hull voxel carving and Signed Distance Field (SDF) optimization—achieving MAE(L,W) = 7.5 mm and MAE(H) = 10.2 mm without depth sensors. 2. Multimodal Physical Constraint Inference: Batches high-resolution RGB item crops into Meta Muse Spark (muse-spark-1.3 / muse-spark-open-1026 on Baseten) to identify each product and infer physical handling rules (`fragility`, `keep_upright`, `stackable`, `weight_g`), guarded by a fail-closed deterministic safety sanitizer. 3. Exact 3D CP-SAT Combinatorial Solver & Interactive 3D CAD Guide: Feeds verified metric dimensions and physical constraints into a Google OR-Tools CP-SAT 3D bin-packing solver. OptiPack evaluates standard warehouse box SKUs to generate 4 Pareto-optimal strategies (Lowest Billable DIM Cost, Maximum Fragile Protection, Highest Volumetric Density, and Custom Carton) with an interactive Three.js 3D step-by-step packing visualizer.