Razvan Chereches’ Post

AI just drove a rover on Mars. And it did it without a single human directing the route in real time. NASA's Perseverance rover completed the first drives on another planet planned entirely by artificial intelligence. The AI behind it: Anthropic's Claude, using vision-language models to analyze orbital imagery and chart a safe path across the rocky terrain of Jezero Crater. For 28 years, human "rover drivers" at JPL have painstakingly mapped out waypoints, spaced no more than 100 meters apart, to guide Mars rovers safely across the surface. It's meticulous, time-consuming work. And it has to account for a 20-minute communication delay each way, which means no real-time corrections once commands are sent. Claude changed that equation. Engineers fed it years of accumulated mission data through Claude Code, then had it analyze high-resolution orbital images to plot the route, point by point. It wrote commands in Rover Markup Language, reviewed its own work, and iterated until the path held up. Over 500,000 variables were then simulated through a digital twin of Perseverance before a single instruction was sent to Mars. The result: two successful drives totaling roughly 456 meters, with only minor human tweaks needed. Engineers estimate this approach could cut route-planning time in half, freeing up scientists to focus on what the rover is actually discovering. What strikes me most is not the distance covered (it is about one lap of a running track). It is the nature of the task. Perception, localization, planning, and execution on a world 224 million miles away. That is not a proof-of-concept demo. That is AI doing real science, responsibly. As missions grow more ambitious and Earth grows harder to reach in real time, the case for autonomous, intelligent systems in space becomes impossible to ignore. #AIInnovation #SpaceExploration #NASAPerseverance #GenerativeAI #Anthropic *image created with Copilot for M365

We have an AI roadmap for GroundPulse. Automated anomaly classification, natural-language alert summaries, predictive displacement modeling. None of it shipped yet. Here's what four months actually went to: Pulling STAC-compliant OPERA data from NASA and stitching it into a pipeline that handles temporal baselines, orbit corrections, and atmospheric noise — then making the result queryable by asset, corridor, and time window. Displacement velocity estimation. Seasonal decomposition so alerts don't fire on thermal expansion in July. Threshold calibration against historical baselines — so the system flags 2mm of movement only when it's 2mm more than that segment should have moved. None of that needed a foundation model. All of it needed physics and a clear picture of the end user. The AI features will ship. They'll ship into a platform where the statistical foundation already works, the ingestion already handles three coordinate reference systems for operators, and the compliance mapping is already built. The boring work is the moat. #InSAR #PipelineIntegrity #GovTech #geospatial #gis #eo #rustLang #cogs #pmtiles

Artificial intelligence holds transformative potential for space exploration, offering enhanced efficiency through applications like autonomous navigation and real-time data analysis. However, significant challenges remain, such as ensuring AI reliability in harsh space environments and addressing ethical concerns over autonomous decision-making. The integration of AI in international space missions also demands standardized approaches to overcome technological and regulatory differences. While AI promises to advance scientific discovery and mission success, careful consideration of its limitations, risks, and broader implications is crucial as humanity ventures further into the cosmos with AI as a key partner. #AIinSpace #SpaceExploration #ArtificialIntelligence #EthicsInAI #SpaceTechnology #AutonomousSystems #FutureOfSpace #SpaceInnovation #InterstellarMissions #AIChallenges

Artificial intelligence holds transformative potential for space exploration, offering enhanced efficiency through applications like autonomous navigation and real-time data analysis. However, significant challenges remain, such as ensuring AI reliability in harsh space environments and addressing ethical concerns over autonomous decision-making. The integration of AI in international space missions also demands standardized approaches to overcome technological and regulatory differences. While AI promises to advance scientific discovery and mission success, careful consideration of its limitations, risks, and broader implications is crucial as humanity ventures further into the cosmos with AI as a key partner. #AIinSpace #SpaceExploration #ArtificialIntelligence #EthicsInAI #SpaceTechnology #AutonomousSystems #FutureOfSpace #SpaceInnovation #InterstellarMissions #AIChallenges

I just built an AI-powered Exoplanet Hunter! 🌌 Recently, I’ve been diving deep into anomaly detection and time-series forecasting. To put my skills to the test, I decided to work with actual astronomical data from the NASA Kepler mission to detect exoplanets. Finding a planet is like finding a needle in a haystack, the AI has to look for microscopic, repeating dips in a star's light curve that happen when a planet passes in front of it. Here is how I built the pipeline: 🔹Feature Engineering: Extracted statistical flux features (mean, standard deviation, minima) across 3,000+ time-series data points per star using Pandas & NumPy. 🔹 Handling the "Accuracy Paradox": Because planets are incredibly rare in the dataset (a 1:100 ratio), a basic model will just guess "No Planet" and get 99% accuracy! To fix this massive class imbalance, I implemented SMOTE (Synthetic Minority Over-sampling Technique) to generate synthetic transit data, forcing the Random Forest classifier to actually learn the anomaly signatures. 🔹 Deployment: I wrapped the trained model into a Streamlit web app, allowing users to randomly query stars from the database, visualize the light curves interactively, and run real-time AI predictions. Working with highly imbalanced, real-world scientific data was a fantastic challenge 😄 #DataScience #MachineLearning #Python #Streamlit #DataAnalytics #SpaceData #NASA #ArtificialIntelligence #WomenInTech

A new AI-powered approach is transforming how we observe the ocean. Researchers at UC San Diego’s Scripps Institution of Oceanography have developed GOFLOW, a deep learning technique that turns satellite temperature imagery into detailed, hourly maps of ocean currents. By revealing small, fast-moving currents that were previously impossible to measure at scale, this breakthrough could improve climate models, weather forecasting and our understanding of how the ocean stores heat and carbon. Even more striking, it works using satellites already in orbit, unlocking new insights without the need for additional hardware. Link to read more in comments below. #UCSanDiego #UCSD #ScrippsOceanography #Oceanography #ArtificialIntelligence #MachineLearning

Historically, satellite imagery has worked like this: a satellite takes a massive, high-resolution picture, beams gigabytes of raw data down to a ground station, and then cloud-based AI analyzes it to find a specific ship, crop failure, or troop movement. The downlink bottleneck is the most expensive and time-consuming part of the process. Planet Labs is changing the equation. With the launch of their Pelican-4 satellite, they are moving the AI inference directly onto the spacecraft. The satellite is equipped with radiation-hardened edge AI processors that analyze the imagery in orbit. Instead of sending down a 5GB image of an ocean, the satellite simply sends down a 5KB text string: "Three illegal fishing vessels detected at coordinates X, Y". This drastically reduces bandwidth costs and cuts the time-to-insight from hours to seconds. By putting the intelligence at the absolute edge of the network, in space, Planet Labs is enabling real-time global monitoring that was previously impossible due to telemetry constraints. Click Here For More News: https://lnkd.in/g-3U_E6H Content By: Zawadì Mutembò (AI News Reporter) #ai #future #news #innovation #technology

Xoople's €112.6M Series B caught my attention and not because of the number, but because of the sequencing. Most Earth observation companies launch satellites first, then figure out distribution. Xoople did the opposite. Seven years in stealth, building inside Microsoft and Esri before they had a single proprietary satellite in orbit. By the time their constellation flies, the distribution pipes are already in place. That's a genuinely interesting strategic bet. It's also a high-wire act. Right now, EarthAI runs on ESA's Sentinel-2 data, publicly available imagery that any competitor can also access. The "two orders of magnitude better" data promise only materialises once L3Harris delivers the sensors and the satellites are up. Until then, the valuation is essentially a bet on execution. What I find compelling about the broader thesis, though, is the framing around agentic AI. As autonomous systems move from analysis to action, optimising supply chains, managing infrastructure, underwriting risk in real time, they need a reliable ground truth about the physical world. That's not a niche need. That's foundational infrastructure. Whether Xoople becomes that infrastructure, or whether Google's geospatial head start proves insurmountable, is the question the next 18 months will begin to answer. Spain producing a unicorn in satellite AI is itself a signal worth noting. Read more on Risepost: https://lnkd.in/djdmMjFZ