DIM-WAM: World-Action Modeling with Diverse Historical Event Memory

A new memory-augmented model called DiM-WAM sharply raises robot-manipulation success rates by retaining diverse historical event information, according to a preprint posted to arXiv on June 26, 2026 [1][2]. The model, described as a world-action architecture, addresses a persistent limitation in robotic control: existing systems rely on short-term history and short-horizon predictions, which proves inadequate for long-horizon tasks where correct execution hinges on earlier observations and overall task progress [2]. DiM-WAM integrates multi-scale historical context, local future dynamics, and global task progress to counter long-term forgetting and poor awareness of the global task state [2]. Its memory module extracts compact visual event information from real observations, updates multiple memory banks through independent similarity-based merging, and reads bank-identity- and time-embedded long-term context to condition video and action denoising [2]. A progress-supervision objective further encourages memory tokens to encode completed historical events, the current task stage, and implications for the remaining task [2]. On the RMBench benchmark, DiM-WAM lifted average success from 28.4 percent with the LingBot-VA baseline to 69.8 percent, surpassing the explicit-memory Mem-0 baseline, which scored 42.0 percent [2]. In four real-world Franka tasks, average stage success improved from 70.7 percent to 91.5 percent, and full-task success rose from 52.5 percent to 80.0 percent [2]. The paper was submitted to arXiv, an open-access repository that hosts preprints across physics, computer science, and related fields and that surpassed two million articles by the end of 2021 [6]. The repository’s submission rate reached roughly 24,000 articles per month as of November 2024 [6]. arXiv also supports community-built tools through its arXivLabs framework, which allows collaborators to develop experimental features such as bibliographic explorers and code finders directly on article pages [5]. The framework, launched in 2020, sets guidelines ensuring that partners share arXiv’s values of openness, community, excellence, and user data privacy [5]. The DiM-WAM preprint page includes links to several such tools, including the Bibliographic Explorer and CORE Recommender [4][5]. The authors have also provided a project page for further details [2].