Repository navigation
Releases
SAGE v1
SAGE_V1
SAGE v1.0.0 is a non-neural, interpretable object recognition system for robotic manipulation: objects are fit as superquadrics via nonlinear least-squares, so every number the system reasons over is a physical quantity (radius, roundness, color) rather than a learned embedding, and category vocabulary is learned online with no gradient descent anywhere in the pipeline. This release ships a YOLO-style API (SAGE.load("SAGE_V2"), .predict(), .train()) with the trained weights (SAGE_V2.json, attached below) auto-downloaded and cached on first use. On held-out YCB-Video data it reaches 89.4% accuracy from a single observation and 95.7% when multiple viewpoints are fused before classifying (a real, statistically significant improvement, p=0.0072) — and to be upfront about the honest tradeoff, a full black-box baseline (PointNet) still beats it on raw accuracy (93.2%), since SAGE's actual contribution is interpretability, not state-of-the-art accuracy.