ASH: New AI System Learns Complex Tasks from Unlabeled Internet Video Without Expert Guidance
Researchers have introduced ASH, an AI agent that teaches itself to complete long-horizon embodied tasks by learning from unlabeled internet video without hand-engineered rewards or expert annotations. The system uses a self-improvement loop in which it builds an Inverse Dynamics Model from its own experience to extract supervision from relevant online video, and employs unsupervised learning to identify and retain key moments as long-term memory. Evaluated over 8-hour sessions on Pokémon Emerald and The Legend of Zelda: The Minish Cap, ASH substantially outperformed all baselines, suggesting self-improving agents may offer a scalable path to long-horizon embodied AI.
ASH (Agents that Self-Hone) is a new agentic framework designed to address a core bottleneck in embodied AI: the need for costly hand-engineered rewards or action-labeled demonstrations. Instead, ASH enters a self-improvement loop — when it gets stuck, it trains an Inverse Dynamics Model (IDM) on its own collected trajectories and uses that model to extract supervisory signals from noisy, unlabeled internet video. An unsupervised mechanism identifies key moments in large-scale video and stores them as long-term memory, enabling the agent to plan across multi-hour horizons. The system was benchmarked on two demanding game environments: Pokémon Emerald, a turn-based RPG, and The Legend of Zelda: The Minish Cap, a real-time action-adventure game. ASH achieved an average of 11.2 out of 12 milestones in Pokémon Emerald and 9.9 out of 12 in Zelda, compared to the strongest baseline's 6.5 and 6.0, respectively, with all baselines — including behavioral cloning, retrieval-augmented, and zero-shot foundation-model approaches — plateauing well before the evaluation ended. The paper was accepted as a workshop contribution at the ICML 2026 Workshop on Scalable Learning and Optimization for Efficient Multimodal AI Agents.
What's missing
The study uses video games as proxies for embodied tasks, and it remains an open question how well ASH's self-improvement loop would transfer to physical robotic environments or tasks with real-world consequences. The paper does not report computational costs, training time, or the scale of internet video required, which are important factors for assessing practical scalability. Additionally, the quality and diversity of the internet video corpus used are not characterized, leaving uncertainty about how sensitive performance is to video source selection.
What different sources said
- arXiv cs.AICenter
ASH: Agents that Self-Hone via Embodied Learning
Related
Gut Bacteria Enzyme Found to Break Down Heat-Processed Food Compounds, Producing Novel Biogenic Amines
Researchers have discovered that an enzyme in common gut bacteria can degrade N-epsilon-carboxymethyllysine (CML), a compound formed during thermal food processing, producing previously unknown biogenic amines. The enzyme, ornithine decarboxylase SpeC from enterobacteria, acts on CML and related modified lysine derivatives through a low-level 'underground' catalytic activity. This finding suggests a previously unrecognized communication axis between thermally processed dietary compounds and gut microbial physiology, with potential implications for host health.
Full-Length Gene Sequencing Reveals Two Distinct Bacterial Communities in Black-Legged Ticks Expanding Into Canada
Researchers used Oxford Nanopore full-length 16S rRNA gene sequencing to characterize the microbiome of Ixodes scapularis black-legged ticks collected in Nova Scotia, Canada, distinguishing between tick-adapted bacteria and environmentally acquired bacteria. The study comes as I. scapularis — the primary vector of Lyme disease — is rapidly expanding northward into Canada due to climate change. The findings suggest that environmentally derived bacteria in tick microbiomes are not mere contamination, which has implications for how tick microbiome data is collected and interpreted across surveillance studies.
Study Identifies Metabolic Link Between Cell Envelope Stress and Biofilm Formation in Bacteria
Researchers have discovered that the metabolite acetyl-CoA directly inhibits enzymes that degrade the bacterial signaling molecule c-di-GMP, connecting cell envelope biosynthesis stress to biofilm formation in Pseudomonas aeruginosa. The study found that sub-inhibitory concentrations of antibiotics targeting early peptidoglycan biosynthesis — but not other antibiotic classes — elevate c-di-GMP levels by reducing phosphodiesterase activity, with acetyl-CoA competing for the enzyme active site. Because the relevant enzyme domain is broadly conserved across bacterial species, this checkpoint mechanism may be widespread and could have implications for understanding antibiotic-induced biofilm responses.