Pruning Attention Layers Slashes Action Expert FLOPs by 94% in Robotics

mathildepapillo · x · 2026-08-05

Based on deep insights into the model interface, developers fine-tuned the MolmoBOT and Pi05 models, retaining only specific attention layers (e.g., L24 and Pi05 L14-16). This drastically reduced the action expert's computation (cutting FLOPs by 94% and 15%, respectively), with no statistically significant difference found in success rates or execution times.

Original post →

More from Embodied

Embodied channel →