--- kind: trend trend_doc_id: 70 granularity: week period_start: '2026-03-02T00:00:00' period_end: '2026-03-09T00:00:00' topics: - robotics - vla - world-models - memory - deployment run_id: materialize-outputs aliases: - recoleta-trend-70 tags: - recoleta/trend - topic/robotics - topic/vla - topic/world-models - topic/memory - topic/deployment language_code: en --- # Robot VLAs move toward deployable systems: on-demand reasoning, memory plugins, and safe world models ## Overview Robot research converged strongly this week. The central question is clear: how do we move VLAs and world models from “they can do it” to “they can do it reliably, efficiently, and in deployment.” One major thread is on-demand reasoning . Many systems no longer assume that a large model should be called at every step, and instead let high-level reasoning appear only at key moments. This both saves compute and better suits long-horizon tasks. Tri-System is a representative example of this idea: it inserts a Critic between a high-level vision-language model and a low-level controller, maintains a fast closed loop during normal execution, and only triggers replanning when progress stalls or anomalies appear. The second thread is that memory capability is starting to be decomposed . The research focus is shifting from “does the model have a memory module” to “what kind of memory does the task actually require.” On one side are benchmark efforts like RoboMME, which evaluate temporal, spatial, object, and procedural memory separately. On the other side are lightweight methods like TempoFit, which try not to modify the backbone and instead extend long-horizon capability mainly through caching mechanisms. ## Clusters ### VLAs shift toward on-demand reasoning and failure recovery The strongest theme this week is pushing vision-language-action models (VLAs) from demo systems toward deployable systems. Methods are no longer trying to do heavy re-reasoning at every step, and instead emphasize on-demand activation, asynchronous scheduling, and failure recovery. The representative work is Tri-System: it uses a visual Critic to monitor execution and only wakes the high-level VLM when a subtask is completed, an incident occurs, or progress stalls, clearly outperforming single-system and dual-system approaches on real long-horizon tasks. #### Representative sources - [Critic in the Loop: A Tri-System VLA Framework for Robust Long-Horizon Manipulation](../Inbox/2026-03-05--critic-in-the-loop-a-tri-system-vla-framework-for-robust-long-horizon-manipulation.md) — Pengfei Yi; Yingjie Ma; Wenjiang Xu; Yanan Hao; Shuai Gan; Wanting Li; … ### Robot memory moves from concept to evaluation and plug-in enhancement Memory is no longer just about “giving the model history.” This week emphasizes two things more: first, clearly measuring what memory is actually needed; second, adding temporal capability in lighter-weight ways. RoboMME splits memory into four categories—temporal, spatial, object, and procedural—and shows that there is no one-size-fits-all solution. TempoFit instead takes a plugin approach, directly reusing layer-wise K/V caches to improve long-horizon manipulation success rates without training. #### Representative sources - [RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies](../Inbox/2026-03-04--robomme-benchmarking-and-understanding-memory-for-robotic-generalist-policies.md) — Yinpei Dai; Hongze Fu; Jayjun Lee; Yuejiang Liu; Haoran Zhang; Jianing Yang; … - [TempoFit: Plug-and-Play Layer-Wise Temporal KV Memory for Long-Horizon Vision-Language-Action Manipulation](../Inbox/2026-03-08--tempofit-plug-and-play-layer-wise-temporal-kv-memory-for-long-horizon-vision-language-action-manipulation.md) — Jun Sun; Boyu Yang; Jiahao Zhang; Ning Ma; Chencheng Wu; Siqing Zhang; … ### World models move toward structured dynamic representations and safety interfaces The focus of world models is shifting noticeably. The priority is no longer generating more realistic video, but learning dynamic representations that are useful for control and connecting those representations to safety monitoring and decision-making. CoWVLA replaces redundant future-frame reconstruction with latent motion chains and reaches 0.956 on LIBERO. Another line of work uses probabilistic world models for runtime anomaly detection, achieving 92.0±6.4% overall accuracy on bimanual failure detection. #### Representative sources - [Chain of World: World Model Thinking in Latent Motion](../Inbox/2026-03-03--chain-of-world-world-model-thinking-in-latent-motion.md) — Fuxiang Yang; Donglin Di; Lulu Tang; Xuancheng Zhang; Lei Fan; Hao Li; … - [Foundational World Models Accurately Detect Bimanual Manipulator Failures](../Inbox/2026-03-07--foundational-world-models-accurately-detect-bimanual-manipulator-failures.md) — Isaac R. Ward; Michelle Ho; Houjun Liu; Aaron Feldman; Joseph Vincent; Liam Kruse; … ### Lightweight adaptation and viewpoint robustness become a deployment patch layer Another major thread in real deployment is changing the model less and adding more interface-layer fixes. AnyCamVLA, without adding demonstrations or fine-tuning the policy, transforms test-time viewpoints back into training-time viewpoints in real time, significantly improving robustness to camera perturbations. Contemporary lightweight adaptation work is also reducing task transfer costs, showing that the community is now treating “how to deploy at low cost” as equally important as “how to improve scores.” #### Representative sources - [AnyCamVLA: Zero-Shot Camera Adaptation for Viewpoint Robust Vision-Language-Action Models](../Inbox/2026-03-06--anycamvla-zero-shot-camera-adaptation-for-viewpoint-robust-vision-language-action-models.md) — Hyeongjun Heo; Seungyeon Woo; Sang Min Kim; Junho Kim; Junho Lee; Yonghyeon Lee; …