Researchers led by Chen Zhao have built InteractVLA, which turns pretrained flow-based vision-language-action policies into tools that an AI agent can steer and inspect, according to a paper on arXiv. In today's agent systems such a policy only takes a text command, so when it fails, the agent can only try again or hand over to a person.
With InteractVLA, the agent can give a target position for the gripper, which guides the policy while it runs without bending its approach. In the other direction, the agent sees the planned actions drawn into the camera images, so it can spot where they go wrong, change the target for the next call or adjust parts of the motion.
Used for rounds of self-improvement, the system reached 99.4% on LIBERO-Plus, 89.4% on RoboCasa365 and 48.62% on RoboDojo, which the authors say are the best results on these benchmarks.



