DuckMan

Using archived MechDog camera images and sonar labels, we taught @BrainChip_inc's AKD1500 to tell a nearby box from a farther one in camera images. It then answered all 96 fresh settled views correctly at measured 3-foot and 7-foot positions, with teaching off and no range reading in the query. All 60 earlier-view checks remained correct. The 110 archived lessons made 742 connection replacements in chip memory. A never-trained control got 21/96 right; disconnecting the native readout removed every supported answer. The early-frozen memory and CPU reference also got 96/96, so the later lessons showed no additional accuracy benefit on these views. This is a small stationary test across four correlated holds of one box, not 96 independent scenes. Real recordings were replayed through the chip. The computer prepares image inputs, supplies archived teaching labels, paces calls and interprets the class scores. The film shows measured chip outputs and memory changes. Neuron highlights are calculated from recorded memory and image inputs. Connections show stored feature selections, not physical wires. Geometry is schematic and timing is edited. Next: change the visual condition and test whether new sonar feedback corrects the answers while earlier knowledge survives.

MechDog x AKD1500 We want MechDog to use sonar to teach its vision about distance, remember useful visual cues, and keep learning as its surroundings change. This experiment tests a building block: learning new relationships between familiar image parts without losing earlier lessons. The eight patches come from our stationary MechDog camera recording. We reduced them to small pixel grids, then rearranged them into controlled test pictures. The earlier demo added synthetic pieces and their relationships. Here, the new lessons add relationships among already learned pieces. After new relationship lessons, BrainChip's AKD1500 completed all 8 new arrangements and still completed all 8 earlier ones. With later lessons disabled, it completed none of the 8 new arrangements. In 4 cases it kept several possible answers open. In 4 inconsistent cases it returned unknown. No complete test picture was taught. This run learned the relationships on chip. It did not use sonar lessons or demonstrate depth prediction, obstacle detection, steering or generalization to ordinary camera scenes. Those remain goals. The computer prepared inputs, supplied labels for the designed relationships, configured each step and clocked a designed completion procedure. CPU controls also pass. The film replays captured chip responses and memory changes. The connecting lines show decoded associations, not physical wires. The layout is schematic and the timing is edited.