From AI that supervises a robot mission to the molecular machinery that copies a genome. Experiments, models, and the evidence behind what they can do.
Saved simulations, with the evidence in view.
14 articles
MissionOS connects AI judgment, human approval, constrained execution, and outcome verification across drones, robots, and a Starship-inspired simulator. A design and development report grounded in saved simulation cases.
MissionOSの管制レイヤーを、横浜でのVLA・WAM配送、移動ロボット、操作修復、Starshipの再計画・診断から整理する。共通の権限・証拠管理と、成功・停止・未検証を分けた試験記録。
The full development story: a continuous flight simulator, AI mission decisions, fault diagnosis, physical-model audits, and the separately registered October 23 integrated mission.
MissionOS separates model judgment, human authorization, constrained execution, and outcome verification, with explicit evidence for each transition.
▶ Video & figuresPredictive stopping improved simulated stacking scores over weak baselines, while strong simple rules and known-case integration runs define the limits of the claim.
Controlled LIBERO experiments separate corrective targeting from stable repair, preserving negative results and the limits of comparisons among GR00T, Cosmos Policy, and VLA-0.
▶ Video & figuresTwo recorded Go2 simulator runs connect approval, navigation, simulated receipt, and return while showing exactly when DeepSeek was—and was not—invoked.
▶ Video & figuresNative VLA and world-model proposals completed a bounded simulated delivery, while separate wind and gust experiments exposed unresolved recovery limits.
▶ Video & figuresLanding-pad studies separate verified waiting and permission renewal from unqualified native forecasts and modest, correlated replay-only decision changes.
Ship-to-shore experiments distinguish route-selection value, camera-based execution, and native-model integration while preserving failed implementation cohorts.
▶ Video & figuresA development simulator connects approved mission scope to live return replanning while distinguishing record integrity, modeled contact, and real vehicle capability.
Across 105,000 simulated observation sequences, acquisition cadence and molecule count support different levels of mechanism discrimination and mixture estimation.
A genome-scale investigation corrected an intuitive explanation, tested finite reusable resources, and reached the limits of public-data validation—without establishing a new biological mechanism.
Factorial simulations separate recycling from binding and loading, while public initiation-rate tables and a clock-scaling control show why curve shape alone cannot validate a human replication mechanism.
No matching articles. Try another keyword or select All research.
Astropome AI is an independent research project exploring governed AI, autonomous systems, and computational biology. These articles bring together the question, the method, the measured result, and the limits of each study.
The original MissionOS collection is based on a pinned snapshot of the public main branch. Dated updates may additionally include reviewed saved-record summaries and media, identified in each article. The biology articles introduce curated summaries of saved computational studies, with links to their published biological context and supplementary evidence.
These are technical research articles, not peer-reviewed publications. Simulator outcomes, synthetic data, and empirical observations are identified within each article. Negative results remain part of the record.
