Choose the control interface
Control width does not determine sequence length. A one-coordinate actuator can have delayed physical consequences; a seven-coordinate command applies its values together. Choose the smallest useful control interface that leaves the intended behavior for Adapt-1 to acquire.
Manipulation through pose targets
Use this pattern when fixed actuation software can execute a requested pose, while Adapt-1 learns how to coordinate the poses.- Measure the object and goal using the declared sensing pipeline, together with robot proprioception.
- Request a bounded sequence of position, orientation, and gripper controls.
- Execute each target through fixed inverse kinematics and joint actuation.
- Record the applied controls and measured successor states.
- Report progress or stable placement, then let the next attempt use retained experience.
Aiming with finite-speed controls
Return rates for the aiming axes and a firing value. The executor applies these over time; it does not turn a target location into an analytical shooting solution. The starting context should describe the current joint state and target-relative geometry. A whole-attempt proposal commits the commands before movement. Contextual refinement can then learn how to adjust the acquired sequence for a different target observation. Define the objective separately from the controls. A hit is a completion outcome; continuous miss quality provides more detailed feedback. Record the actual projectile and actuator dynamics as part of the environment contract. Start with acquisition, then refinement.Control delayed physical effects
A carriage can move a suspended load through a single velocity request. Swing is a consequence of carriage motion, so the observation and outcome contract need to account for what continues after a command changes. Supply measured position, velocity, and relevant load motion. Define success over a settling interval, including speed or sway tolerances when the task requires them. Reaching the target at one instant is a different objective from remaining settled there. Keep command holds, neutral padding, and the assessment interval fixed across candidate comparisons. Removing a command must follow the declared time convention. See structural stages for comparing shortened and reordered executions.Short execution windows
Use short blocks when the environment needs another observation before the full task is over. Each block remains an ordered numeric sequence.1
Observe at the boundary
Measure the current environment state. Request a sequence with the configured horizon and the current goal.
2
Execute the window
Apply commands with the fixed cadence. Stop if the environment terminates; record only the executed prefix.
3
Report the measured consequence
During learning, return the actual controls, successor states, and attributable outcome. Reconcile the acknowledgement before the next learning request.
4
Continue from the resulting state
If the task is still active, observe again and request the next window. Reset only at the environment’s real episode boundary.
episode_credit mechanism name alone does not establish credit across every block in an application episode; provide the outcome contract supported by your integration.
The acquisition guide gives the observation body. For a named discrete action after each observation, use Sequential Discovery.
Compose maneuvers through continuous state
Use this pattern when your application supplies a sequence of objectives and each acquired maneuver must accept the state produced by its predecessor.
Supplying an objective order is part of the application contract. It does not establish that the learner discovered that decomposition. Avoid training each maneuver only from an artificial stationary start if it must later accept a moving handoff.
Keep each maneuver’s Domain, selected base, context encoding, and executor version together. Separate Domains make these contracts inspectable; composition itself runs through the continuous environment state.
Correct an acquired behavior
Preserve a useful base, then learn a bounded correction from the incoming state. Include the observations that can explain the needed change: target-relative position, velocity, changed actuator strength when measured, or another pre-action condition. Use the contextual refinement workflow to execute complete candidate comparisons. When earlier behavior must remain reliable, include original conditions in the practice mix and compare the corrected policy with its unchanged base on the same reserved cases. Check three properties separately:
The retained-use guide covers the API calls and the manifest needed for reuse.
Choose feedback that describes the outcome
Use a binary completion value when that is the only dependable measurement. Use measured stage progress or a graded physical error when your application can supply it consistently. More informative feedback can make initial acquisition easier; it still describes the consequence of an attempt rather than providing the successful controls. Keep the success rule separate from the learning score. A distance score may improve while the required stopping or release condition still fails. Intermediate outcomes must come from actual measurements; if you only measure a final outcome, omitstep_outcomes.