Researchers led by Xuji Zhang have built RecallFlow, a flow-based robot policy for tasks in which the robot has to act on things it can no longer see, according to a paper on arXiv. Existing policies add the observation history as an extra input but still start generating actions from random noise, so they have to find the relevant target in a long history.

RecallFlow builds its starting distribution from episodic memory instead. A few recall slots link what the robot remembers to its current view, so the start reflects where remembered targets are now, and each slot becomes one part of the distribution. Generation therefore begins close to the intended action, and the rest comes from the current view alone, so the policy runs almost as cheaply as one without memory.

On the public MIKASA-Robo benchmark for memory it reached 80.4% average success, and it beat the strongest baseline by 20.3 points on the authors' own set of memory tasks and by 18.8 points on four tasks with a real robot.

Source: arXiv Robotics (cs.RO)

Background

Xuji Zhang

Robotics researcher

Robotics researcher.