• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Announcement

Memory tokens are claimed to boost robot policies on memory-critical tasks

A post describes pretraining an encoder to compress vision and proprioception histories, then adding its tokens to policies via FiLM.

3 Sources, 3h ago, first seen 3h ago

TLDR

A post describes a policy-agnostic approach to robot memory: pretrain an encoder to compress long histories of vision and proprioception into compact tokens, integrate them into existing policies through FiLM, then fine-tune them together. The post claims even a handful of tokens can deliver big gains on memory-critical tasks.

Combined views

—

3 Sources, first seen 3h ago

— likes— comments— saves— reposts

Combined views

—

3 Sources, first seen 3h ago

— likes— comments— saves— reposts
Image: @YilinWu11 on X

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

3 Sources

Yilin Wu@YilinWu11What to remember in the past to act—spatial, temporal or motion details? We often assume what’s useful, but that’s hard to know in advance. Meet IMPRINT 🧠: retain rich memory first, then learn to use it for control. Compact latents keep inference fast as history grows. 🧵 1/103h
Guanya Shi@GuanyaShiA lightweight, policy-agnostic way to boost existing robot policies’ memory and in-context learning! The key idea: pretrain a memory encoder to compress long histories of vision and proprioception into compact “memory tokens,” then integrate them into policies via FiLM and jointly fine-tune. Even a handful of memory tokens can deliver big gains on memory-critical tasks!2h
Gokul Swamy@g_k_swamyRT @YilinWu11: What to remember in the past to act—spatial, temporal or motion details? We often assume what’s useful, but that’s hard to…1h
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    3 Sources

    Yilin Wu@YilinWu11What to remember in the past to act—spatial, temporal or motion details? We often assume what’s useful, but that’s hard to know in advance. Meet IMPRINT 🧠: retain rich memory first, then learn to use it for control. Compact latents keep inference fast as history grows. 🧵 1/103h
    Guanya Shi@GuanyaShiA lightweight, policy-agnostic way to boost existing robot policies’ memory and in-context learning! The key idea: pretrain a memory encoder to compress long histories of vision and proprioception into compact “memory tokens,” then integrate them into policies via FiLM and jointly fine-tune. Even a handful of memory tokens can deliver big gains on memory-critical tasks!2h
    Gokul Swamy@g_k_swamyRT @YilinWu11: What to remember in the past to act—spatial, temporal or motion details? We often assume what’s useful, but that’s hard to…1h
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet