We’re taking a look at the trends that matter most for leaders navigating the next 12 to 36 months.

We organize trends into three categories based on their maturity and impact timeline:

Game Changers: Trends reshaping entire industries right now. These affect how businesses operate, how value is created, and how competitive advantage is built.

Foundational Breakthroughs: Scientific and engineering advances unlocking new possibilities. These create strategic optionality for the decade ahead.

Weak Signals: Early indicators of transformations that will dominate the late 2020s. These require positioning now, even if mainstream adoption is years away.

Today, we’re diving into one of the Foundational Breakthrough trends: Omnimodal Intelligence.

What is it?

Multimodal AI (text + image + audio) was only the beginning. Omnimodal systems can combine vision, language, spatial data, code, simulation, physics, and robotic action, enabling AI to understand not just our digital worlds but the physical world around us.

A good metaphor for this is the human brain. We don’t have one brain for seeing, one for hearing etc. We have one integrated brain that considers all our senses, all at once. This is the foundation for robotics, AR, autonomous systems, and digital environments that understand us as richly as we understand them.

Where we’re seeing it emerge

Every sense fused into physical action. Figure’s Helix 02 folds vision from head and palm cameras, tactile sensing fine enough to register forces as small as three grams, proprioception, and natural language into a single neural network that controls an entire humanoid robot.

Rather than bolting a locomotion controller onto a separate manipulation system, one model lets the robot walk, balance, and manipulate as a single continuous system, running entirely from onboard sensors with no human intervention. In one demonstration, it completed a continuous four-minute task made up of 61 separate actions, described as the longest-horizon autonomous task a humanoid has performed to date. It signals where omnimodal is heading: perception fused tightly enough that a machine can operate inside the physical world it senses.

One model that sees, listens, and moves a whole body. In July 2026, Google DeepMind introduced Gemini Robotics 2, a family of models that takes in live video, audio and language and turns them into action. Its reasoning model can talk with people, track its own progress on a task by watching it, and plan multi-step work, and in one demonstration it paused a humanoid when a person walked nearby and resumed only once the area was clear. The action model now directs an entire humanoid, not just its hands, and an on-device version runs on the robot’s own computer and can be adapted to a new robot with a few hours of training. Omnimodal perception is moving onto the machine itself, where the decisions happen.

3D worlds built for machines to train in. World Labs’ Marble generates persistent 3D worlds from text, images, video, or coarse layouts, and it does so as a multimodal world model rather than a single purpose image tool. What makes Marble a genuine signal of omnimodal progress is where its output goes next: exported as Gaussian splats or collider meshes, these AI generated environments are already being pulled directly into NVIDIA Isaac Sim to train robots, cutting scene setup from weeks to hours. Marble is a bridge between generative AI and the physical systems that have to operate in the real world.

A robot brain that imagines before it acts. NVIDIA’s next humanoid model, GR00T N2, previewed in March 2026, is built as a “world action model.” Rather than mapping what a robot sees straight to a movement, it first predicts how the scene will change, then plans the actions that get it to the goal. NVIDIA says this helps robots succeed at new tasks in new environments more than twice as often as leading vision-language-action models, and N2 currently ranks first on the MolmoSpaces and RoboArena benchmarks for generalist robot policies. It is slated to be available by the end of the year.

The world’s robots learn in are changing too. Cosmos 3, which NVIDIA released as an open model in June 2026, handles text, images, video and robot actions in one system, folding scene understanding, world generation and action simulation together. NVIDIA calls it the first open omni-model for physical AI reasoning and action. Together they show omnimodal perception moving from understanding a scene to controlling a body inside it.

Omnimodal perception as the eyes and ears of AI agents. Most agent systems today still stitch together separate models for vision, speech, and language, losing time and context every time information passes from one model to the next. NVIDIA’s Nemotron 3 Nano Omni, released in April 2026, folds text, image, audio, video, and document understanding into a single open model built for exactly this handoff problem, delivering up to 9x the throughput of comparable open omni models. It is built for computer use agents, document intelligence, and audio and video reasoning. Palantir and Foxconn are among the companies already adopting it, and others such as Oracle and Docusign are evaluating it, an early sign that omnimodal perception is moving from research benchmark to production agent stack.

Why it matters

For years, “multimodal” meant a model that could look at an image and describe it or transcribe audio into text. Omnimodal is a different ambition: systems that hold vision, sound, language, spatial structure, and physical dynamics in one coherent understanding, the way a person walking through a room does without thinking about it.

That shift changes what AI can do in the places where decisions actually get made:

Decisions made in context. An omnimodal system can read voice tone, facial cues, documents, dashboards, and live video at once, connecting a spike in sensor data to the line in a written report that explains it and picking up both what is said in a meeting and what is left implied. Leaders get fewer blind spots because the signals no longer sit in separate tools that never talk to each other.

Holistic diagnosis. In medicine, one model can weigh imaging, genomics, lab results, and a patient’s speech patterns together, flagging anomalies that surface across modalities rather than in any single test. That gives clinicians integrated insight instead of four disconnected readouts, pointing toward earlier detection and better outcomes.

Just-in-time operations. On the factory floor, omnimodal perception fuses visual inspection, IoT sensor streams, and maintenance logs into real-time anomaly detection across machines and whole facilities, with hands-free guidance delivered through voice and augmented visuals. The payoff is reduced downtime, safer environments, and predictive intervention before failure instead of after.

Immersive learning. In training and education, a system can learn from how a student speaks, writes, gestures, and interacts, then adapt instruction across AR/VR, video, and dialogue and drop the learner into experiential simulations. Because instruction meets each person where they actually learn, retention climbs and training moves faster.

The through line is the same in every case: when perception stops being fragmented, a system can act on the full picture instead of a slice. For enterprise leaders and founders building in this space, the strategic question is where perception stops being a bottleneck. Teams working on robotics, spatial computing, simulation, healthcare, industrial operations, or agentic AI are the ones who will feel this shift first, and the infrastructure choices they make now, which world models to build on, which perception stacks to bet on, will shape what they can ship over the next several years.