Research

What would change an AI’s mind?

An intelligent system should be able to state what it expects, identify the assumptions behind that expectation, and choose an experiment that could prove it wrong. We are building toward that capability in physical systems.

From a plausible future to a testable explanation.

A model predicts what follows from its assumptions. An experiment can challenge those assumptions. Vivacity’s research centers on the connection: structured state, controlled interventions, measurable outcomes, and a record of how the evidence changes the next decision.

Industrial robotics is our first proving ground. A failed placement has concrete geometry, contact conditions, and a controller to examine. It also exposes a difficult boundary: a result inside a simulator remains a result about that simulator until it is checked against physical measurements.

Vivacity visual study of a flow field around a boundary
A visual study of physical dynamics. Image fidelity alone does not establish model validity.

Questions guiding the work

The following is a research agenda. Methods and evaluation criteria describe what we intend to investigate; they are not claims of published results.

Which experiment should come next?

An action that completes a task and an action that teaches us about the task can be different choices.

A slipped grasp might be explained by low friction, an incorrect mass estimate, or a poor approach pose. Repeating the same movement may leave those explanations indistinguishable. We want agents to identify interventions under which competing models predict different outcomes, then weigh the value of that information against the cost of obtaining it.

Approach

Study parameter sweeps, Bayesian experimental design, and uncertainty-aware search under fixed trial budgets. Compare with random search and expert-designed tests before adding more complex agents.

What would count as progress

Can the method reduce uncertainty or uncover independently reproducible failures with fewer trials on held-out scenarios?

What should change when a model is wrong?

A discrepancy is evidence. It does not identify its own cause.

A predicted trajectory can disagree with measurement because of a wrong parameter, missing dynamics, sensor timing, or an inadequate observation model. Fitting a residual can reduce error while concealing the mechanism. The challenge is to distinguish what can be calibrated from what requires a different explanation, and to recognize when available observations cannot tell the difference.

Approach

Investigate system identification and constrained residual dynamics. Preserve known mechanics, estimate uncertain parameters, and evaluate corrections on motions and conditions excluded from fitting.

What would count as progress

Does a correction improve predictions on unseen conditions, with uncertainty that reflects where it stops being reliable?

When are two trials really comparable?

Reproducibility begins below the visible scene.

Identical joint positions do not necessarily imply identical experiments. Controller memory, contact state, random seeds, solver settings, and sensor timing can all change the result. A useful branch must preserve the conditions relevant to the question. Across simulators, the meaning of a contact or a reset may differ; a shared interface must expose those differences.

Approach

Define snapshot contracts for each backend, pin model and evaluator versions, and test replay and branch isolation. Keep numerical outputs, learned predictions, and physical observations distinguishable in the evidence record.

What would count as progress

Can another engineer recover the initial conditions, reproduce the result within stated tolerances, and trace a disagreement to its source?

Can an agent revise the explanation itself?

Sometimes every hypothesis in the original set is inadequate.

Choosing among fixed parameters is only part of scientific investigation. An agent may need to propose a missing mechanism, derive an observable consequence, and design a test that could reject it. Our longer-term ambition is to connect this process to executable physical models, while keeping explanation quality separate from fluent language.

Approach

Begin with bounded scientific environments where hypotheses make quantitative predictions. Compare model revision with fixed-model fitting, and retain negative results and failed explanations.

What would count as progress

Does the proposed mechanism improve held-out prediction and survive a discriminating experiment, rather than merely explain the data already seen?

The result must survive the test.

We want to measure whether an experimentation layer improves engineering decisions over a well-built simulation script. That requires fixed budgets, meaningful baselines, unseen conditions, and an honest accounting of failed trials.

The live workspace is an early, bounded robotics demonstration. Autonomous failure diagnosis, hardware calibration, and scientific model discovery remain development and research objectives. We do not claim published benchmarks or general autonomous discovery.

Current scope and limits

Foundations we learn from

External work that informs these questions. These are not Vivacity publications, integrations, or endorsements.

TD-MPC2

Planning and control through learned latent dynamics.

BoxingGym

Evaluating experimental design and model discovery in bounded environments.

MuJoCo computation

The numerical and contact semantics an experiment interface must respect.

Working on the boundary between models and measurements?

Talk research with us