Physical Time as a Missing Dimension of Robot Policy Memory Dynamics

Key points

• Existing robot policy memories typically evolve with observation sequences, implicitly treating each update as the same amount of physical time.1
• Yet identical observation histories can span different durations and require different actions, exposing physical time as a missing dimension of policy memory.1
• We evaluate TFP with Diffusion Policy and π{0.5} through controlled comparisons of memory designs.1
• Real-robot tasks requiring timing or memory, together with LIBERO evaluations, demonstrate benefits in physical and general manipulation.1

Sources (2)

• [1]Physical Time as a Missing Dimension of Robot Policy Memory Dynamics
  Hugging Face Daily Papers · Oct 14, 12:00 AM

  We introduce Temporal Flow Policy (TFP), which learns memory dynamics to retain past events and evolve the memory state over measured time intervals.

  Existing robot policy memories typically evolve with observation sequences, implicitly treating each update as the same amount of physical time.

• [2]Physical Time as a Missing Dimension of Robot Policy Memory Dynamics
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 05:24 PM · same content

Extractive summary: sentences quoted from the sources.