IMPRINT: memory tokens give robot policies in-context memory with minimal overhead

GuanyaShi · x · 2026-10-15

• IMPRINT pretrains a memory encoder to compress long vision + proprioception histories into compact "memory tokens," integrated into existing policies via FiLM and jointly fine-tuned.
• The approach is policy-agnostic: even a handful of memory tokens yields large gains on memory-critical tasks while keeping inference fast as history grows.
• Motivation: what past details matter for control (spatial, temporal, motion) is hard to know in advance, so retain rich memory first and learn to use it.

Original post →

More from Embodied

• Rhoda's tactile-sensor robot retrieves objects blindfolded, sensing texture and heat — DhruvBatra_ · 2026-10-15
• Valeo and KE:SAI launch open-source initiative for generalist autonomous driving AI — abursuc · 2026-10-15
• ASENA: A Coding Agent Writes and Fixes Programs On-the-Fly to Drive a G1 Humanoid — Lianhuiq · 2026-10-15
• Bittensor SN44 bets on generative world models as a simulation engine for robotics — richdotca · 2026-10-15
• Words to World: A Benchmark Testing LLMs on Closed-Loop Real-World Physical Control — Xinnan Dai · 2026-10-15
• Researcher muses on building a RoboParty robot from scratch in his own machine shop — andrew_n_carr · 2026-10-15

Embodied channel →