Reference actions v0.1 · Recorded evidence

Camera, command,
response.

Watch what the camera can see, what the controller asks for, and how the arm moves.

Two learned controllers · recorded simulation

Eight attempts. Every outcome retained.

Two fixed learned candidates each attempt the same two development starts and two destinations. These are native simulation recordings with no teacher intervention. They are not unseen confirmation trials or a live robot.

Native simulation camera recording
Loading recorded camera evidence…
0.00 s

RGB display is sampled at 2.5 Hz from the original 25 Hz camera. The slider and measurements retain every 50 Hz motor decision. Frame time is shown separately; display sampling does not change the controller's observations.

Object estimate
—
Source observation age
—
Controller
—
Decision authority
—

The target is not the measurement.

Measured at decisionIssued targetApplied at interval end

Recorded actuator tracesMeasured position, issued target and rate-limited applied target. Current values appear below.

When the object estimate is available

Available Missing. Availability says an estimate was returned; it does not certify its accuracy.

Fast predictions still need reliable actions.

All eight attempts complete their recorded control intervals, but neither candidate completes a strict placement: 0/4 absolute and 0/4 reference-relative. These small, reused development cases do not establish a population success rate or a comparison against the original fixed controller.

The learner sees causal camera estimates, explicit missing-value masks, joint measurements and previous applied references. Native object and contact truth is used only for independent scoring. There is no teacher fallback. The two earlier infrastructure batches contain sixteen refusals before learned control; they remain in the research report.

All plotted measurements retain their original values. Issued targets and final rate-limited servo targets occur at different times; the gripper value is one finger's slide position, not total jaw opening. Capture and arrival timestamps are simulation clocks. Playback and inference speed do not establish hardware response time.

Independent readback binds every displayed attempt. The research report separates training, open-loop replay, native failures and their limitations. Both models remain unqualified research candidates.