Physical Time as a Missing Dimension of Robot Policy Memory Dynamics
Changing the elapsed-time input systematically advanced or delayed the robot’s actions.
Temporal Flow Policy retains past events in a memory state that evolves over measured time intervals. In controlled comparisons using Diffusion Policy and π₀.₅, it achieved higher average success than recurrent and state-space memories given the same intervals. Memory readouts recovered time since an event and motion information after occlusion. The authors also report benefits on real-robot tasks requiring timing or memory and on LIBERO manipulation evaluations.
HF Daily Papers' note
Temporal Flow Policy retains past events in a memory state that evolves over measured time intervals. In controlled comparisons using Diffusion Policy and π₀.₅, it achieved higher average success than recurrent and state-space memories given the same intervals. Memory readouts recovered time since an event and motion information after occlusion. The authors also report benefits on real-robot tasks requiring timing or memory and on LIBERO manipulation evaluations.
HF Daily Papers' note
score 5