Explore the hypothesis space.
A wider search can move beyond the nearest minimum. The candidate determines which region of the model is tested next.
We’re building virtual labs where AI agents can run controlled experiments in physical systems.
When a robot misses a grasp, repeating the task is only one option. Change the friction. Test another approach. Ask which result would distinguish one explanation from another.
An agent needs to decide what to test and how to interpret the result. We study both: searching for informative interventions and using observations to constrain a physical model.
A wider search can move beyond the nearest minimum. The candidate determines which region of the model is tested next.
Successive observations refine a damping estimate. The residual reveals the disagreement that remains after each update.
Analytic illustrations of candidate search and system identification. Observations are synthetic; these are not Vivacity benchmark results.
Physical reasoning needs somewhere to meet consequences. Vivacity is developing an experimental workflow around simulation: establish a starting state, change a condition, run the trial, and examine the result.
The goal is a useful loop between hypotheses and evidence—with the conditions of each experiment kept explicit.
Explore the approachA convincing simulation can still rest on the wrong assumptions. Our research asks how agents can recognize that boundary, design a discriminating test, and revise the model.
Our research agendaSolar magnetoconvection visualization · NASA / Timothy Sandstrom
External scientific visualization, not a Vivacity result.
Robotics gives us a concrete place to begin: geometry, contact, control, and failures we can observe.
Our early workspace is a bounded robotics demonstration. The arm beside you is a live, kinematic pick-and-place visualization. Select it to explore the workspace.
What’s available todayWe are working toward agents that do more than predict an outcome. They should choose useful tests, learn from disagreement, and preserve enough context for another engineer to check the result.
Have a robotics task with a failure you need to understand? Help shape the tools around a real engineering question.
Talk to us