Geometry Conditioning in 0.8B Hybrid Model for Robotic Manipulation

Conditioning recurrent decay gates on geometric increments achieves 28.9% success on LIBERO-Spatial tasks using a 0.8B hybrid model with 6.2M trainable parameters. State-only relative-coordinate policies demonstrated higher robustness to frame relabeling than visual policies, which failed under 5 cm object displacements.