A movement needs a meaningful endpoint#

A request to move a particular distance or reach a visible destination needs more than a gait start. The October 15 locomotion work gives MetaHuman a goal-bearing movement interface, connects it to measured progress and preserves the result of finishing that same gait.

The application and P4 support have recorded installations. Reliable model selection, real camera calibration and physical goal accuracy remain separate, unfinished qualification steps.

What a locomotion goal describes#

Goal	Evidence relevant to progress
Distance	Measured displacement in an identified reference frame
Relative turn	Signed orientation change from the starting reference
Approach	A specified relationship to an identified target
Follow	An ongoing target relationship with an ending condition
Visual endpoint	Image evidence evaluated against the selected completion criteria

The model selects the goal, target, tolerance and relevant visual behavior alongside gait controls. Speed percentages are actuation settings; they are not metres per second. Finishing a number of gait cycles or running for a duration does not establish arrival at a real-world destination.

MetaHuman’s advertised contract now requires these goals for locomotion instead of bare Walk, Run, Turn, Crab or Crawl starts. Body Control’s manual gait primitives remain available through their existing controls, and finite gestures such as Bow retain ordinary command handling.

How progress returns to the task#

Core’s existing Active Task owner executes the selected movement and tracks progress. The Environment Bridge carries sensor evidence and the movement’s start, update and Finish results. Sensor updates can resume that execution without another Intent or Task-selection call. Visual and follow goals use the existing behavior evaluator and its selected observation cadence.

Endpoint evidence, requesting Finish, acknowledgment of settings and the original gait’s actual terminal result remain distinct. A spoken claim of arrival does not substitute for any of them.

The owner-approved handling of missing measurements is explicit: report the goal incomplete without starting if its required measurement is unavailable. If that measurement disappears during movement, request Finish on the same gait and retain the outcome for the planner. The system does not substitute timed travel, invent a new target or automatically restart the movement.

What the sensors can and cannot establish#

Position, orientation, camera evidence and target measurements retain their source, reference, time, units and uncertainty. Reading an old sample again does not make it fresh. Camera availability and orientation availability are separate; missing metric camera calibration does not erase valid orientation evidence.

Metric camera estimates need real intrinsics, a measured camera-to-body transform and a measured reference on the ground plane. That estimator tracks the plane; it is not general depth sensing, inertial odometry or room navigation. A metric target needs an explicit ground-contact point in retained evidence. The bottom of a detection box is not automatically a floor contact.

An image-relative relationship can remain image-relative rather than being presented as metres. Visual-only goals use camera evidence without requiring metric calibration or YOLO. A slow visual inference is also different from a camera that has stopped supplying evidence. Commanded foot geometry remains a prediction, not measured ground displacement.

Sideways correction and installation#

The earlier side-crab primitive could not correct sideways overshoot with only forward and turn controls. Signed lateral control now extends the same P4 motion owner, protocol and Core controller.

The October 15 record verifies installation of 0.8.51-p4-lateral, capability negotiation with Body Control, unchanged saved servo calibration, and application activation with goal and position/yaw/camera evidence reaching the Bridge. The subsequent 0.8.52 Wave installation preserves that lateral work. This is installation and data-delivery evidence, not measured sideways correction accuracy.

The models still need qualification#

In an initial eight-case replay without action dispatch, the existing 0.8B Task adapter failed all six locomotion cases while retaining the finite-Bow and discussion controls. The 9B Planner selected valid measured goals in four explicit distance/turn cases, but fabricated visual target evidence in another case. Neither result qualifies complete autonomous movement.

Task training for the new contract was authorized; the reviewed record does not establish a qualified replacement adapter. The separate 9B autonomy/persona diagnostics also retained failures under serving tests.

A destination does not establish a traversable route. Obstacle avoidance, drop-off handling, actual camera calibration and physical goal completion remain open. This update documents the machinery for selecting and evaluating movement without presenting it as demonstrated navigation.

This report uses the Robot Bible’s October 15 locomotion specification, current goal normalization/evaluation source and its focused contract tests. The source snapshot manifest identifies those local records. No robot command was sent for this website update.