EvoMem-VLA adds state-evolution memory for long-horizon robot manipulation, hitting 80.7% on RMBench
zhenjun_zhao · x · 2026-10-08
EvoMem-VLA: memory for long-horizon manipulation
VLA models typically rely only on current observations, losing task-relevant evidence once it leaves view. EvoMem-VLA builds state-evolution memory that explicitly encodes observed changes between historical states instead of isolated snapshots.
- Conditional delta tokenization: ordered frame pairs become directional, source-conditioned delta tokens preserving interaction-outcome evidence
- Task-adaptive routing: a shared VLM backbone routes normal long-horizon tasks to direct actions, multi-stage tasks to a subtask route
- Results: a single jointly trained policy scores 80.7% on RMBench, 82.0% on RoboMME, and 83.8% across four real-world tasks on two robot embodiments
More from Embodied
- Hierarchy-GBP accelerates factor graph inference via coarse abstraction and recovery, SOTA BA runtimes — zhenjun_zhao · 2026-10-08
- Clear Oakley Meta Vanguard smart glasses earn Z87 jobsite safety certification — armand_ruiz · 2026-10-08
- DeepMind's Raia Hadsell on continual learning, navigation and real-world robotics opportunities at RAAIS — nathanbenaich · 2026-10-08
- Robot fights got too dangerous, so organizers put swords on the bots — cixliv · 2026-10-08
- From single-task to full-workflow: why physical AI robots unlock a 10x larger market — Rewkang · 2026-10-08
- VibeBuddy is a $59 desk gadget that watches your coding agents and speaks up when they need you — juntao · 2026-10-08