How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Design note, 2026-08-05, written before the change was measured. Follows the rendezvous note, which argued that a grasp needs an arrival; this fixes what the encounter becomes once the arrival has to be produced by servoing rather than by a script.
| capture radius, measured | < 1 mm |
| arm's steady-state error following a moving script | 7.3 mm |
| separation a servo reaches, no script to chase | 0.002 mm |
Seven-fold apart, and the 7.3 mm is the IK controller's, not the schedule's — it was 7.5 mm at 2 mm/step and 7.7 at 2.5, and no speed removes it. A script-following approach cannot produce a capture in this scene. The measurement
The obvious shortcut is to grasp before the episode starts, the way --preroll
brings the arm to the start line, and record only the carry. That is not
available, and the reason is the paper's rather than the simulator's.
The commit window is one dispense-length either side of the arrival. If the arrival happens before step 0, the window has no anchor inside the episode, and every commit sits in the riding tail — which is exactly the regime the window was introduced to exclude, where a constant-velocity model absorbs the motion and the single-entity arm eventually matches the relational one.
So the episode must contain: a still target, an approach, the taking hold, and the carry. In that order, observed.
Under a script the encounter gave the body's positions at each step, and the arm chased them. Under a servo it gives the direction instead:
before the grasp the arm goes to the object, wherever the object is
at the grasp the gripper closes; this is the arrival
after the grasp the arm moves along the drawn axis, on the burst schedule
The drawn azimuth still decides which way the carry runs, and it is still drawn per seed, so encounters remain distinguishable from one another. What is dropped is the pretence that the body's position was known in advance — it never was, since the arm could not reach it.
Nothing about the relation changes: the target is perfectly still until a body
takes hold, and then it rides. That is what capture_displacement has always
said and what the physics has now been shown to do.
contact_arrivals anchors the commit window on a body crossing
interaction_radius — 12 mm in this scene. The grasp happens under 1 mm. At
roughly 3 mm/step those are about four steps apart, so the anchor and the
capture are not the same event, and arm D cannot act until it has three riding
steps after the capture.
Whether that leaves a usable overlap is a measurement, not a choice. It is
recorded here in advance that the three quantities involved —
interaction_radius, the commit window's half-width, and min_carriage_run —
are not to be adjusted against each other after seeing the result. The window
is fixed on the structure of the action, the run on the collision equilibrium,
and the radius is a property of the scene to be measured. If they do not overlap,
that is a finding about this task, and the honest response is the same one the
onset question got: write down what the design cannot do.
What may change on measured grounds is the encounter — how fast the carry runs and how long the bursts are — since those are the encounter's own parameters and have never been fixed. Any such change is stated before the run that tests it.