AnyBox Achieves 36-Point Leap in Zero-Shot Box Detection for Warehouse Robots

Robotic systems handling warehouse inventory need to identify and grasp boxes regardless of clutter, occlusion, or stacking. Existing approaches stumble: model-based pipelines demand CAD files for each product variant, while generic methods struggle with the featureless surfaces and symmetry of cardboard boxes.

A new technique called AnyBox sidesteps these limitations by exploiting what boxes are—geometric primitives with predictable structure. The framework recovers all nine dimensions of box identity (6D pose plus 3D size) from a single RGB-D image, without any object-specific training or CAD data.

How It Works#

The method begins with a generic box template and iteratively refines both orientation and scale. At each step, it compares where the template projects onto the observed scene versus where the actual object mask appears, using that gap to narrow down true dimensions via binary search.

Two design choices make this practical in chaotic warehouse settings:

• Depth-consistency filtering eliminates phantom solutions that arise from box symmetry, where multiple orientations might fit the same 2D silhouette
• Early stopping replaces exhaustive search with a single geometric calculation once the solution is sufficiently constrained

Performance and Impact#

On public benchmarks and proprietary warehouse datasets, AnyBox nearly doubles previous results, gaining up to 36 percentage points in detection average precision and matching performance of systems that have access to ground-truth CAD models. The gains carry forward to downstream tasks: robotic success on a cluttered box-shelving challenge jumped by 28%.

The framework's efficiency and zero-shot nature make it practical for logistics operations where inventory constantly shifts and full CAD coverage remains impractical.