Researchers led by Zheyuan Liu have built State-Aware Fallback Imitation (SAFI), a way to make a vision-language-action policy stop a prohibited task even when it takes over partway through, according to a paper on arXiv. The authors found that policies trained to refuse a task at the start could still finish it from a later state, and might drop a held object if they stopped.

SAFI labels prohibited demonstrations with commands that hold the arm still and keep the gripper as it is, and trains the policy on these alongside its permitted tasks. In tests with OpenVLA-OFT on LIBERO tasks, success at prohibited tasks after a takeover fell to 1.3%, with no recorded physical incidents, while permitted tasks kept a high success rate.

Source: arXiv Robotics (cs.RO)

Background

Zheyuan Liu

Robotics researcher

Robotics researcher.