Perception & Sensing

Seeing the bin, not just the object.

Structured-light depth, RGB surface classification, and fingertip tactile — fused into a single persistent scene graph, not three isolated readings. The system builds a 3D model of the entire workspace, not just the object currently in the grasp envelope, so objects that leave the field-of-view are not treated as new unknowns when they reappear.

Sensor Architecture

Three modalities, one consistent world model

Structured-light depth gives a dense 3D point cloud of scene geometry regardless of surface colour or ambient lighting variation — the two failure modes that defeat RGB-only vision in factory environments. The RGB camera runs in parallel, adding surface material classification, label reading, and part-type identification. Tactile sensors in the fingertip pads close the loop on contact: they confirm actual surface contact, measure grip force distribution, and detect slippage before the part is dropped.

All three feeds merge in a scene graph running on embedded compute at the head unit. The graph persists objects across partial occlusion events — when a bin neighbour moves in front of the target object and then away, the system does not lose track and attempt a new full-scene classification. This reduces the re-identification overhead on every grasp cycle.

The Bin-Picking Challenge

Why vision alone is not enough

2D camera-based grasp planning fails in four common factory scenarios: mutual occlusion between adjacent parts, specular reflections off metal surfaces, ambient lighting variation across shifts, and bins of geometrically identical objects. Each one is a routine condition on a real factory floor. None is an edge case.

Atom's approach: plan the grasp on geometry first, classify the object second. The depth sensor finds a surface with viable contact geometry regardless of what the object is. The RGB sensor identifies it during the reach motion. Tactile feedback at the fingertips confirms contact quality before the lift. Tasks that fail camera-only planning succeed because the grasp decision does not depend on colour or texture recognition alone.