Robotics · 2026-10-06
EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation
Yuheng Na, Zhide Zhong, Junjie He, Junfeng Li, Haodong Yan, Jiaan Wang, Jiaguan Zhu, Yangyang Zheng, Tianyu Huang, Haoang Li
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
EvoMem-VLA targets long-horizon, memory-dependent robot manipulation in VLA by adding state-evolution memory that explicitly encodes observed changes between historical states. It uses conditional delta tokenization for ordered frame-pair change tokens and a shared VLM backbone with task-adaptive routing (direct vs subtask route). The authors report success rates of 80.7% (RMBench), 82.0% (RoboMME), and 83.8% on real-world tasks.
Why it matters
The authors argue that reactive VLA policies lose task-relevant evidence after it leaves view and that storing only sparse snapshots leaves ambiguity about what changed. By linking retained visual states with learned change evidence (directional, source-conditioned delta tokens) and routing for multi-stage tasks, EvoMem-VLA aims to preserve interaction-progress evidence beyond isolated observations for both simulation and real-world robot control.
Method
- Construct state-evolution memory by interleaving initial, recent, and selected keyframe observations with conditional delta tokens that encode ordered source–target changes.
- Use a shared VLM backbone for action prediction with task-adaptive routing: direct route for normal long-horizon tasks and subtask route that generates an executable subtask as additional context for multi-stage tasks.
- Train with a two-stage optimization: learn delta tokenizer components with frozen DINOv3, then freeze them and train the VLA (action head, delta projection, and keyframe head) on demonstrations.
Limitation
The paper text states that for real-world evaluation, they use two robot embodiments with a separately trained policy per embodiment, and per-task success is pooled across embodiments; no further limitations are explicitly stated.
Abstract (from arXiv)
Most vision-language-action (VLA) models rely on current observations and lose task-relevant evidence once it leaves view, limiting performance on long-horizon, memory-dependent tasks. Existing efforts incorporate compressed historical features or sparse visual keyframes. However, isolated snapshots can leave the policy uncertain about what changed during past interactions and which action should follow. To overcome this limitation, we propose EvoMem-VLA, which constructs state-evolution memory by explicitly encoding and retaining observed changes between historical states. These change representations preserve evidence of interaction outcomes, allowing the policy to track task progress beyond isolated snapshots. Specifically, we introduce conditional delta tokenization to encode ordered frame pairs into directional, source-conditioned delta tokens, each associated with its corresponding state evidence. A shared VLM backbone supports task-adaptive routing: normal long-horizon tasks follow a direct action route, whereas multi-stage tasks use a subtask route that generates an executable subtask as an additional input for action generation. With a single jointly trained policy for each simulation benchmark, EvoMem-VLA achieves success rates of 80.7\% on RMBench, 82.0\% on RoboMME and 83.8\% across four real-world tasks spanning two robot embodiments. These results represent substantial improvements over the previous state of the art in all three evaluation settings.
Related papers
- MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation
- T$^2$Mem: Learning Test-Time Memory for Robotics
- Fast-VLA: You Only Think Once for Robot Manipulation
- Plan First, Act Later: Streaming Reasoning for Executable Memory in Robot Control
- Memory Without Curation: Learning Multi-Stream Latent Memory for Robot Policies
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-04 · Indexed by Pub-AI: 2026-10-06 UTC
EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation
Preprint posted: 2026-10-04 · Original record ↗
- arxiv
- 2610.05418
Explicit source versions (1)
- v1 · 2026-10-04
- Preprint available
Source date: 2026-10-04 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Most vision-language-action (VLA) models rely on current observations and lose task-relevant evidence once it leaves view, limiting performance on long-horizon, memory-dependent tasks. Existing efforts incorporate compressed historical features or sparse visual keyframes. However, isolated snapshots can leave the policy uncertain about what changed during past interactions and which action should follow. To overcome this limitation, we propose EvoMem-VLA, which constructs state-evolution memory by explicitly encoding and retaining observed changes between historical states. These change representations preserve evidence of interaction outcomes, allowing the policy to track task progress beyond isolated snapshots. Specifically, we introduce conditional delta tokenization to encode ordered frame pairs into directional, source-conditioned delta tokens, each associated with its corresponding state evidence. A shared VLM backbone supports task-adaptive routing: normal long-horizon tasks follow a direct action route, whereas multi-stage tasks use a subtask route that generates an executable subtask as an additional input for action generation. With a single jointly trained policy for each simulation benchmark, EvoMem-VLA achieves success rates of 80.7\% on RMBench, 82.0\% on RoboMME and 83.8\% across four real-world tasks spanning two robot embodiments. These results represent substantial improvements over the previous state of the art in all three evaluation settings.