Product guide

Working concepts.

A practical overview of the ideas shaping Vivacity. This is not final API documentation.

An early robotics workspace, with a broader experimentation system in development.

Overview

Virtual labs for investigating physical decisions.

Vivacity is developing an experimentation runtime for AI agents, beginning with industrial robotics. The aim is to connect a physical question to a configured scenario, controlled changes, repeatable execution, and evidence that can guide the next trial.

Current scope

The live workspace exposes bounded robot tasks with physical settings, recorded trials, measurements, and an execution log. The broader architecture described here is in development; it is not a released general-purpose API.

World models

A world model predicts. An agent still needs a system for acting on that prediction.

A learned world model estimates how a system might evolve. A physics simulator executes specified dynamics. Our proposed runtime organizes the experiments around those models: which assumptions changed, what ran, what was measured, and which result should inform the next decision. Backend interoperability is a development goal, with semantics that must be validated separately for each tool.

Wt+At→E→Wt+1

World state plus action is routed through an execution process to produce a candidate next state. This notation is conceptual.

Environments

Start with a precisely defined scenario.

For manipulation, a scenario includes robot geometry, joint limits, object placement, contact parameters, and a controller. A reproducible baseline also needs simulator configuration, seeds, controller memory, and any internal state that affects the result. Saving a picture or joint pose alone is insufficient.

  • ContinuityState survives beyond one inference turn.
  • Constraint surfaceThe environment defines the boundary of admissible change.
  • LineageEvery candidate state retains its causal origin.

Actions

An action is a proposed state transition.

An experiment specifies the starting conditions, the intervention, and the criteria used to interpret the result. For example: keep the approach trajectory fixed, vary object mass, and measure placement error. A useful action interface must retain units, coordinate frames, joint constraints, and controller-specific limits. These are proposed interface requirements, not public SDK guarantees.

Observations

Keep a measurement connected to its source.

A rendered frame, a simulated contact force, and a hardware sensor reading provide different kinds of evidence. Their source, time, units, and uncertainty should remain visible. Comparing predicted and measured behavior only makes sense when the conditions and observation processes are aligned.

Scenario state

The persistent variables, constraints, relations, and time index.

Observation

A task-conditioned projection for the next decision.

Execution

Route each transition to the physics it requires.

Closed-form dynamics may call for an exact solver. Contact-rich mechanics may require a physics engine. Open environments may benefit from learned rollouts. Private systems may use domain logic. The runtime should preserve a common causal trace while keeping each executor's evidence intact.

EXACTPHYSICSLEARNEDCUSTOM

No third-party integrations are claimed here. Execution paths are research directions and partnership topics.

Status and limits

What you can explore today.

  • Robotics workspaceA bounded browser demonstration of robot tasks, configurable conditions, recorded trials, and measurements. Controller limits are described in the workspace.
  • In developmentA broader agent interface, validated cross-backend experiments, and systematic evaluation. No public general-purpose SDK is documented here.
  • Research objectivesAutonomous diagnosis, informative experiment selection, hardware calibration, and scientific model discovery. No published performance gains are claimed.
  • Simulation boundaryA successful simulated trial does not establish hardware performance or certify a robot safe.