Google Research Details 10 New AI Initiatives Spanning Edge Chips and Planetary Models
At its Research@ event, Google introduced an experimental Earth AI agent API, Edge AI hardware with Synaptics, document verification via Deep Review, and medical AI milestones.
At its Research@ event in Mountain View, Google Research revealed ten research initiatives highlighting advances in agentic workflows, domain-specific foundation models, edge silicon, and scientific computing. The announcements focus on giving technical teams and domain specialists direct access to systems designed for complex numerical, geospatial, and biological tasks.
Planetary Scale Models and Crisis Forecasting
Google expanded early access to Google Earth AI, introducing three primary avenues for users to interact with its geospatial capabilities. Alongside deeper analytical tools in Google Earth and a planetary prediction engine for risk assessments, Google launched an experimental agent-to-agent API. This interface lets developers programmatically tap into Geospatial Reasoning agent skills, including satellite imagery analysis, routing functions, and demographic embeddings.
The system runs on more than 70 geospatial tools, incorporating foundation models like the Remote Sensing Foundation Model (RSFM) and the Population Dynamics Foundation Model (PDFM), validated in a new Earth AI Technical Report. It also integrates Google Dataset Search, connecting users to over 100 million datasets, and TimesFM for forecasting. In parallel, Google.org committed $10 million in crisis resilience grants, including $2 million to The Milken Institute for core infrastructure projects.
In disaster modeling, Google presented Groundsource, a technique that leverages Gemini to transform public crisis information into structured historical data. The approach produced a dataset cataloging over 2.6 million flood incidents, enabling the training of models that predict urban flash floods up to 24 hours ahead. This complements Google's riverine flood prediction system, which provides forecasts up to seven days ahead across 150 countries.
Scientific Agents, Document Review, and Tabular Intelligence
To accelerate academic inquiry, Google demonstrated a multi-agent orchestrator that replicates published research methods to address new problems. In one fluid mechanics trial, the orchestrator evaluated inverse design models to propose urinary catheter structures that minimized bacterial growth in simulations. This builds on other multi-agent efforts such as Co-Scientist, developed with Google DeepMind, which aided Stanford University researchers in identifying repurposed therapies for liver fibrosis.
For document validation, Google opened public access to Deep Review. The agentic system examines technical papers, checking factual grounding and internal logic. Prior to this release, an earlier iteration helped evaluate more than 16,000 conference submissions in computer science. The tool is now available to inspect documents such as financial statements and legal agreements.
Google is also incorporating its TimesFM and TabFM foundation models into Gemini for business. These architectures process raw time-series feeds and structured spreadsheets without task-specific training, allowing users within Cloud BigQuery to project operational figures, including inventory demand and clinical bed occupancy.
Edge Hardware, Healthcare Models, and Prototyping
On the hardware front, Google announced that Synaptics will reveal an upcoming low-power Edge AI chip co-designed with Google's open-source intellectual property. The chip follows the launch of the Coralboard earlier this year, targeting on-device processing for smart glasses, connected cameras, and industrial sensors.
In healthcare, Google reported that its open MedGemma model has surpassed 10 million downloads among 65,000 innovators. Researchers are equipping MedGemma with tool-use capabilities to query clinical records, alongside experimental support for ultrasound, mammography, and electrocardiograms. Concurrently, a clinical study published in The Lancet examined AMIE at Beth Israel Deaconess Medical Center, where 98 patients consulted the conversational model prior to urgent care visits, with attending clinicians finding the generated clinical summaries useful in 75% of cases.
Google also showcased interface experiments, including generative UI modules for middle and high school STEM education, and Play with Putty, a shared natural-language workspace for multiplayer UI prototyping currently waitlisted on Google Labs.
What it means for developers
Google's latest announcements emphasize specialized APIs over generic text completion. The experimental agent-to-agent Earth AI API allows engineers to automate spatial pipelines, integrating satellite imagery parsing directly into external agent workflows without managing raw satellite telemetry.
For teams working with numerical data, the availability of zero-shot models like TimesFM and TabFM in BigQuery eliminates the need to maintain bespoke regression pipelines for standard operational tables. In healthcare, MedGemma's expanding tool integration provides a documented base for building applications that can interact with patient record systems while processing multimodal diagnostic inputs.
As organizations build multi-agent systems and test specialized architectures across various providers, developers can try top AI models cheaply through one API at https://apixoai.online to benchmark inference performance and token expenditures across providers like Google, Anthropic, and OpenAI.
Source: Research@ 2026: 10 ways we're amplifying human ingenuity — Google Research. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
