Preprint reports weight edit improves robot policy's instruction following

Researchers report in a preprint that directly editing a robot policy's weights increased successful task completions from 27 to 106 across 128 held-out seeds. The policy had released a cube before reaching a bowl because it misread which object the instruction referred to. The method transfers the effect of a clearer instruction into the model's weights without gradients, training data, or changing the original instruction. A released tool reproduced the repair on new held-out seeds, raising success from 8/64 to 36/64. Edits targeting perception, a missing grasp, contact physics, or the gripper threshold did not help. The authors say the approach works best when a known-good input exists to transfer.