Research
AI research explained: what a study found, how it was tested and what its limits mean outside the lab.
AI news, explained simply
Today
OpenAI disbands its stand-alone robotics team
OpenAI is moving robotics researchers into its main research and Applied groups, saying its general-purpose models can now help control robots.
CHIME report shows 93% of health systems now use AI
The 2026 Digital Health Most Wired survey finds AI embedded in clinical or operational processes at 93.2% of participating facilities.
Google opens two datasets used in WeatherNext 3
SatWeave provides hourly satellite mosaics, while PARDIG maps precipitation every 15 minutes. Google says both datasets are openly available for research and development.
CrowdMAPS lets crowds label multiple robot actions per state
A web tool lets remote annotators add up to ten alternative actions for each robot state, boosting success rates by about 35% on real tasks.
XInfer boosts robot fleet efficiency by up to 3.75×
The new XInfer serving system lets hundreds of VLA‑powered robots run more tasks while using fewer GPUs.
Study shows AI leaderboard rankings can be gamed
Researchers find that preference‑based model rankings can be boosted by up to 45 spots using simple manipulation tricks.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
Yesterday
Natera launches SWARM AI system to cut lab ticket response time
The new AI tool built on Amazon Bedrock AgentCore speeds first responses by 130× and automates 96% of support tickets.
May Mobility plans 10,000 autonomous VW ID. Buzz robotaxis by 2031
The company will start delivering adapted ID. Buzz vans in 2028, aiming for over 10,000 on its platform.
Google researcher’s AI “discovery” of Maya site was already documented
Pietro Miotti’s claim of finding a new Maya settlement was debunked when archaeologists noted the site was published in a 2025 journal.
Qvyon launches with $8 million to test autonomous vehicles and robots
Qvyon Technologies offers Mimic Core, a SaaS platform that turns real‑world sensor logs into virtual test scenarios, now available with a 14‑day trial.
Penguin AI survey finds 62% give up care after prior‑auth delays
A November 2026 survey of 500 U.S. adults shows most patients abandon treatment after insurance prior‑authorization hurdles.
Wednesday, November 11
Quectel launches SP900BF-AP module with 100 TOPS AI compute
The new smart module uses Qualcomm's Dragonwing IQ-9075 chip to run large AI models on robots and edge gateways.
Excelra data will support Boltz drug-discovery models
Boltz will use curated GOSTAR measurements to train models that screen drug candidates for properties such as solubility and permeability.
Robo-MLT reports 82% success on a lab robotics task
Researchers describe a robot system that monitors laboratory workflows as it works and can replan after an operator removes a sample. It was tested in two tasks.
RoboACE reports higher robot success on LIBERO-PRO
Researchers describe a system that revises both robot practice challenges and execution routines. Its frozen deployment reached 74% success on LIBERO-PRO.
Tuesday, November 10
Oura to buy Counsel Health for $625 million
The smart-ring maker plans to add AI-supported primary care to its health services. The deal is expected to close before the end of 2027.
AMD and Cornelis broaden partnership for scientific HPC and AI
The two firms will jointly help research groups build systems with AMD EPYC CPUs, Instinct GPUs and Cornelis networking, easing design and deployment.
Pseudo-colour ECG images boost AI heart‑risk detection and explanations
Adding a clinically designed colour map to ECG images lifted AUC from 0.72 to over 0.9 and made model explanations more stable in a UK‑Biobank epilepsy cohort.
GUARD steers robot policies away from risky actions
A new framework uses anomaly scores from a policy’s training data to guide robot actions. The authors report gains in simulated and real-world manipulation tasks.
ReSkill boosts robot task success from 30% to over 80% in simulation
The new selective self‑improvement framework raises average success on ten RoboDojo tasks to 86.2% and reaches 77.5% on real‑world trials.
Monday, November 9
DongA Science’s AiR NOVA AI tutor named CES 2027 Innovation Awards honoree
The Korean media firm’s AI tutor for K‑12 students, built on 40 years of science journalism, will launch worldwide in January 2027.
NVIDIA unveils GR00T 2 and Isaac Lab 3.0 at CoRL
At the Conference on Robot Learning, NVIDIA introduced a new world‑action model and simulation tools that let robots practice delicate assembly tasks in virtual worlds.
Pierre Fabre and XtalPi start AI‑driven drug discovery partnership
The French pharma group and Hong Kong AI firm will use XtalPi’s AI and robotics to design small‑molecule candidates for cancer and other hard‑to‑treat diseases.
Robot policy method coordinates obstacle avoidance across actions
Researchers report 75% collision-free task success in real-robot tests using guidance that coordinates obstacle avoidance across a robot’s action sequence.
VICTR improves robot task success using demonstration progress
A research team reports that VICTR raised success on unseen real-world manipulation tasks to 23.8%, up from 6.2% for a fine-tuned baseline.
PhysioWave-T aims to cut biosignal model costs
Researchers introduce wavelet-based models for EEG, ECG and muscle signals, reporting lower compute costs and improved results on selected tests.
A2-MEAS reports results across robot simulation and household tasks
The arXiv paper introduces a multi-agent robot-learning system and reports strong benchmark results, including improved recovery after failures in coffee preparation.
FFR’s OmniMind model tops RoboColiseum leaderboard
FF EAI Robotics says its OmniMind AI ranked No. 1 on the RoboColiseum simulation leaderboard on Nov 6, scoring 0.8698.
Sunday, November 8
Saturday, November 7
Friday, November 6
PharmAgents2 reports three experimentally checked drug findings
The AI drug-discovery system proposed an inhibitor scaffold, a JAK1 strategy and a new protein target; human experiments later supported the findings.
Deep Principle releases open-weight Hydrogen-1.0 for science
The model is available by download, API and through Mira. Deep Principle says it improved on several scientific evaluations, though details are limited.
MessyNav lets robots clear paths without task-specific training
The framework uses a vision-language model to decide whether to move, pick up or avoid obstacles during navigation.
Robot video codec uses motion data to cut uplink bandwidth
A preprint reports lower video bit rates and higher robot-task success by using joint states and camera data to predict what a robot camera will see.
Thursday, November 5
Axiom turns the Stacks Project into machine-checked Lean code
Axiom says its 16-million-line formalization makes a major algebraic geometry reference reusable in proofs checked by Lean.
Researchers use AI to study dolphin communication
Denise Herzing’s team is using AI to find patterns in dolphin sounds, but researchers say understanding their meaning—and holding a conversation—remains a distant goal.
Agility Robotics to go public in $465 million SPAC deal
The humanoid robot maker plans a public listing and says it has $300 million in planned Digit 5 contracts, subject to performance targets.
Study Finds Gaps in LLM Code Verification
An arXiv paper separates verifier acceptance from proof-to-spec binding and real-world correctness, finding large gaps across three code benchmarks.
A learned goal metric lifts planning success in four benchmarks
A new arXiv paper reports that ranking states by time to a goal raised average planning success from 32.7% to 63.0% across four benchmarks.
Researchers generate 6,575 datasets to test training side effects
An arXiv paper describes a pipeline for finding ordinary-looking training data that unexpectedly changes language model behavior across tasks.
