Skip to main content
Machina adds continuous-control sequence learning to Adapt-1 Preview. Give Adapt-1 an observation, a permitted control interface, and measured outcome feedback. It can acquire coordinated executions from its own attempts, revise them, and retain them for later use. This is Adapt-1 operating in a trajectory and procedural-control regime. The same contextual and temporal learning system now learns executable sequences. Your application supplies the sensors, actuator mapping, and environment that turn those sequences into action.

Configure your first control task

Define numeric inputs and controls, then create a trajectory Domain.

Choose an execution pattern

Choose command timing, observation boundaries, and how acquired routines fit together.

​
Where this could fit

Machina could support task-specific routines inside a robot, a machine, or an interactive environment. Possible integrations include: These are application directions. Current control evidence comes from simulations with task-specific policies; performance on a new environment or physical machine needs separate evaluation. The application ideas page also covers Adapt-1 workflows that use predictions, policy choices, or retained evidence.

​
What your application sends and receives

A command is one row of control values applied together. A sequence orders those rows over time. Its horizon, control width, and command duration are separate choices: eight two-coordinate commands contain sixteen scalar outputs, while their physical duration comes from your executor.

​
Learn through execution

1

Acquire a sequence

Request controls, execute them, and return what actually happened. Later attempts can revise retained executions, including a useful prefix followed by a different continuation. Start with sequence acquisition.
2

Test structural changes

Compare acquired executions on repeatable contexts. Test removing spans or changing the order of control changes, then carry the selected execution forward. Use selection, reduction, and ordering.
3

Refine for the current context

Preserve the acquired base and learn bounded corrections from pre-action observations. Use contextual refinement.
4

Reuse or continue learning

Execute retained behavior with learning stopped, or resume compatible practice. Keep the same Domain identity, base, and executor contract. See retained use and recovery.
These are stages you orchestrate through the API. Choose the stages your task needs; a structural stage may keep its input unchanged, and refinement should be evaluated against the acquired base.

​
Keep the learning boundary clear

For example, fixed inverse kinematics can translate an end-effector target into joint targets. The sequence that grasps, carries, and releases an object remains something to acquire. The execution patterns show how to define this boundary for different applications.

​
Choose when observations can change the controls

Low-level controllers can still respond during a committed sequence. That does not create a new Adapt-1 proposal inside the sequence. See execution windows for the integration loop.
Frozen learning can still produce context-dependent controls. A fixed base repeats the same acquired sequence. A frozen contextual policy uses unchanged learned state to produce corrections for the current observation. Reconnecting to that state is a separate persistence check.

​
Connect the right API

The numeric workflow uses POST /domains/{domain_id}/trajectory/{operation} under the Adapt-1 production API. Its configuration and observation bodies are specific to the selected mechanism. Start with Schemas and configuration for authentication, exact request bodies, and the shared Python helper. The guides use one small two-actuator example throughout, with application hooks for your environment.