A robot is loading a plate into a microwave. It detects a sealed drink can on the plate and removes it first. As the robot begins to close the door, a hand moves into its path, and the robot stops.
The robot is running CRIS-0, a causal intelligence system from the San Diego company Aether AI. UC San Diego assistant professor Biwei Huang founded the company in 2026 after twelve years of research in causal discovery and causal AI. On October 8, Aether AI released footage from CRIS-0’s first public demonstration, showing the company’s causal world model and causal agent operating together on a physical robot.
From predicting patterns to taking action
Most AI models learn to predict what comes next. But a robot also has to predict what its own actions will change.
Prediction can capture patterns in data without identifying what drives an outcome or how an intervention will change the state of the world. In text and software, a wrong step can often be undone. A robot often cannot. Moving an object, opening a door, or pouring a drink changes the conditions for every step that follows. For Aether AI, causality means modeling how interventions change outcomes, so a system can reason about what may happen under different actions before choosing one.
Aether AI is building the infrastructure for causal intelligence. Its first application is Physical AI, where the company develops the system that decides what a robot does next.
Robotics startups raised more than $18 billion in the first half of 2026, and funding rounds and acquisition deals for world model startups have topped $10 billion this year. Aether AI is taking a different approach: rather than relying on correlations alone, it is building models designed to learn how actions cause changes in the world and use that understanding to guide what a robot does next.
“For the past two years, robot hardware has moved faster than the intelligence behind it,” said Prof. Biwei Huang. “I believe the brain will pull ahead next. We’re building causal intelligence that goes beyond pattern matching to reason about the underlying cause-and-effect relationships that shape what happens next.”
How CRIS-0 reasons before it acts
In a continuous loop, CRIS-0 estimates how an action may affect a task, acts, observes the result, and replans when conditions change.
CRIS-0 maintains an explicit causal state, a compact representation of the conditions that matter at each stage. During a grasp, for example, this can include whether an object is securely held. The system breaks long tasks into stages with observable completion conditions, checking the outcome of each stage before moving on. This allows it to correct local failures before they affect the rest of the task.
The reasoning loop centers on a causal world model and a causal agent. The causal world model predicts how an action will change that state and what physical outcome may follow, and a causal agent uses those predictions to plan and revise the task. Aether AI first tested earlier iterations of both in settings where mistakes can be undone. As of September 2026, its causal world model, CausalWM, ranks first on TriWorldBench and first in the robot track of PAI-Bench, two public benchmarks for world models. Its agent framework, RSIAgent, surpasses frontier closed-source models on OSWorld 2.0 and Agents’ Last Exam, two benchmarks for computer-use agents, without additional training. CRIS-0 brings these capabilities together on a physical robot for the first time.
When the world changes mid-task
When a result differs from expectation, CRIS-0 identifies which part of the task was affected. It may retry the action, return to an earlier stage, or revise the plan. If those attempts fail, the system is designed to request assistance and stop.
Aether AI’s internal evaluation tested this loop across several tasks. In a coffee bean pouring and grinding task, CRIS-0 recovered from deliberate disruptions, such as the coffee bag being moved mid-task, and replanned in an average of two seconds. In a microwave scenario, it stopped or adjusted its plan within an average of 0.2 seconds after detecting a safety risk. In the personalized pick-and-place task built around ambiguous requests, it chose and moved the intended object with a success rate of over 90%. In another part of the demo, it prepared a drink for a user back from a morning run, choosing a low-sugar functional beverage that suited her training.
Across these tests, CRIS-0 used the same core loop for every task and disruption. In each case, the loop had to notice that the world had stopped matching the system’s assumptions, work out what changed, and adjust the next move. Aether AI sees that consistency as a foundation for the same reasoning loop to generalize across different application domains and scenarios.
Scaling beyond robotics
The company started with robots because robots punish mistakes. A robot can spill a drink, damage an object, or create a collision risk, so a system that misses a mismatch fails quickly and visibly.
In addition to robots, Aether AI plans to extend the same reasoning pattern to forecasting and scientific discovery, adapting the relevant state, tools, and completion criteria to each. For now, CRIS-0’s evidence comes from these initial robotics demonstrations. Learn more about the company’s work at https://aetherlabs.ai/.
