Project Machina

Project Machina uses Adapt-1 to acquire and retain task-specific actions through interaction. Choose a setup below, define what your application can observe and execute, then connect measured outcomes to the learner. Machina is not an LLM wrapper around a pretrained robot policy. The relevant output is a structured action, parameter, command block, or action sequence that the surrounding application executes. Generated language is not required in the control loop.

​

Start with a working contract

The numeric trajectory guides use one two-actuator example throughout. They cover the production URL, authentication, exact configuration bodies, array shapes, execution hooks, and stage handoffs.

​

The operating boundary

Machina separates the learner from the system that senses and acts. An authored controller is not automatically leakage. A fixed actuator or inverse-kinematics layer can execute a learned target when the example is about acquiring those targets. A supplied route is not appropriate when the withheld capability is route acquisition.

​

Choose the API path

Choose from the output your application needs to execute. The numeric trajectory workflow uses /trajectory/configure, /trajectory/propose, and /trajectory/observe. Its payloads are separate from generic Domain query and feedback payloads. Use the schema for the mechanism you configure.

​

Setup examples

​

Example 1: Image-goal manipulation with a fixed executor

Use this pattern when the learner should acquire task-level movement while conventional software handles low-level actuation. The application loop is:

1. Reset to a new layout.
2. Record only the image-derived and proprioceptive state available before movement.
3. Ask Adapt-1 for the complete end-effector sequence.
4. Convert each target through fixed inverse kinematics and execute it.
5. Measure pickup and placement consequences.
6. Return the declared outcome and retain the resulting learner state.

The fixed executor may translate targets into joint motion, but it must not choose the task-solving target sequence. This keeps the withheld capability at the level Adapt-1 is meant to acquire. Start with Sequence acquisition, replacing the example’s state and action dimensions with your measured interface. If actions are absolute spatial targets, define an explicit coordinate transform before reuse in a new layout; normalized actuator commands and absolute target coordinates need different adapters.

​

Example 2: Whole-sequence differential-drive control

Use this pattern when a complete movement plan can be proposed before motion starts. Adapt-1 emits all sixty values before movement. The executor does not ask for another decision while the platform is moving. This setup is useful when the intended unit of learning is a complete movement program. To make the controller reactive, shorten the plan into blocks and re-observe between them. Use the configuration guide to declare the fourteen state coordinates, two action coordinates, horizon of thirty, and goal mapping. The wheel-speed scale and 0.4-second duration belong to the executor.

​

Example 3: Chunked continuous flight control

Use command blocks when a whole episode is too long for one open-loop proposal but a decision on every simulator tick is unnecessary. The loop is:

1. Observe the current flight state.
2. Request one four-second command block.
3. Execute the full block without replanning inside it.
4. Return the measured trajectory consequence.
5. Request another block if the episode remains active.

Use the trajectory observation contract for numeric command blocks. With eight commands, return nine measured states and eight actually executed action rows. Re-observe before requesting the next block. Keep the same learner Domain and application episode identity across blocks; this alone does not add a separate episode-wide delayed-credit mechanism.

​

Example 4: Retained base with contextual correction

Use this pattern when an acquired behavior mostly works and a measured change calls for a bounded adjustment. A correction loop can:

1. Confirm the preserved parent in its original condition.
2. Observe the changed condition before action.
3. Request a bounded residual.
4. Compose the parent and residual deterministically.
5. Execute the corrected behavior.
6. Return the measured consequence.
7. Retain the correction separately from the parent.

When old-condition preservation matters, include original conditions in the declared practice mix. For later frozen use, keep parent identity, correction identity, composition rule, and context representation fixed. The correction should remain a residual. If it can silently replace the parent, the setup no longer isolates refinement around retained behavior. Follow Contextual refinement for the context matrix, candidate evaluation requests, complete comparison feedback, and base-plus-offset composition.

​

Example 5: One-action persistent-world control

Use this pattern when the same world continues across decisions and every action changes the next observation. Keep one episode identity until the real reset boundary. Every record should contain the ordered step, pre-action state, selected and executed action, next state, consequence, and terminal flag. The action primitives are authored public capabilities. The successful route, action order, and objective strategy are not supplied. Use Sequential Discovery to configure this discrete decision workflow. Its action schema and feedback differ from the numeric trajectory API.

​

Example 6: Staged repertoire and composition

Use separate stages when generating behavior, selecting among candidates, composing a sequence, and learning corrections should remain auditable. At every handoff, keep the source learner identity, state version, candidate identifier, complete structured output, decoder version, qualification cases, and deterministic import rule. Not every task needs every stage. Skip selection when only one useful candidate exists. Skip ordering when one result already describes the complete action. Skip refinement when the parent does not have a bounded, measurable residual problem. Follow Structural stages for paginated record export, separate stage Domains, nomination and confirmation, and transferring selected_execution unchanged.

​

Turn an example into your integration

​

Choose the decision cadence

​

Define what Adapt-1 must learn

Write one sentence:

Given observable state X, return action or sequence Y to improve measured consequence Z over horizon H.

Then list everything supplied by the harness. Sensor calibration, actuator mapping, inverse kinematics, motor control, safety rules, and physics may be valid infrastructure. A successful trajectory, correct route, target-derived oracle value, or hidden evaluator answer is not valid when it supplies the capability being tested.

​

Preserve the application record

For every decision or attempt, keep:

• stable run, episode, step, trial, and event identifiers;
• the exact pre-action observation;
• the complete Adapt-1 response;
• whether Adapt-1, application logic, a controller, or a fallback selected the executed operation;
• the exact decoded operation sent to the environment;
• the measured after-state and native consequence;
• the feedback request and acknowledged response; and
• learner, adapter, executor, simulator, and outcome revisions.

See Monitor adaptation for the complete interaction record.

​

Choose the state boundary

Use Choose a learning setup to distinguish:

• empty-state acquisition during the run;
• a separate acquisition phase followed by frozen use;
• a separately acquired state that continues adapting; and
• a frozen read of retained state.

A new simulator episode does not automatically mean a new learning run. Keep the same learner scope across compatible episodes when accumulated state is intended to carry forward.

​

Freeze before evaluation

Reserve new contexts before acquisition. During frozen use, stop learner writes and exploration, use unseen episode IDs, preserve abstentions, and verify that task-specific state remains unchanged. Keep any application fallback separate from the Adapt-1 result.

​

Continue into the API guides