Why Does ACT's Temporal Integration Weight Past Actions More Heavily Than Current Ones (w_i = exp(-m * i))?

• AB

  All Blues

  5/3/25, 4:30 AM

  Recently, while studying the temporal integration part in the ACT paper, I have a question: The paper states that the weight for past predicted actions is calculated as w_i = exp(-m * i). Why does this formulation give higher weights to past predicted actions compared to current ones?

  0 likes 0 likes Liked Like

  ---

  2
  Posts

  99
  Views

  Reply

  ---

• T

  Tom

  5/5/25, 4:37 PM

  Early-stage motion predictions exhibit higher accuracy, while later stages accumulate errors through recursive deviation. For instance, Time Step 1 may achieve 95% precision, but by Time Step 100, residual errors compound exponentially. Overweighting these error-prone later predictions exacerbates divergence.

  0 likes 0 likes Liked Like

  ---