How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Run 2026-08-05, and it ended the design. Written 2026-08-04 on a machine with no GPU; the iteration it warned about at stage 1 duly happened, and the corrections are recorded in place below and in the linked results.
What it returned. The scene does produce a capture — 60 of 60 cells, so the failure the header below anticipated is not the one that occurred. What occurred instead: Paper 002's mode operator lands 0.957 to 1.000 of physical cells against the relational arm's 0.174 to 0.583. The arm takes 22 steps at the median to come to rest and can hold the block for 54, so the intermittency that makes a relation necessary does not fit inside a carry. Result · Closure
The procedure below is still correct and is kept runnable, since the design reopens for any carrier that can stop inside the time it holds an object.
As written 2026-08-04, kept verbatim. This is the run that could end the design, and that is not a figure of speech. Every Paper 003 result so far is arithmetic: the cell computes the block's motion from a formula and writes it into the simulator's command. Nothing yet shows that the lift scene produces a capture rather than a collision. Preregistration, What the calibration pilot must produce.
A capture needs the target perfectly still before the arrival — if it drifts, its own history carries information and the single-entity arm has something to learn from, which is the failure that killed carriage.
The device is already in the record, as a failure note from the stopping probe:
an open gripper straddles the block — the frame point reached 0.3 mm from its centre without moving it.
For a push probe that was useless. For capture it is exactly right. So:
gripper OPEN, approach block perfectly still, jaws straddling it
gripper CLOSES on arrival takes hold at the observed separation
gripper CLOSED, burst motion the block rides
--grasp schedules that on the observed separation rather than on a step
count, so the block is taken hold of when the arm actually reaches it, not when
the script says it should have. --grasp-radius is the physical capture radius;
arm D does not receive it and estimates it from the observed onset.
Same VESSL workspace and Isaac image as Paper 002; see the Paper 002 runbook for creation. Then:
cd /workspace/research-os && git pull
Nothing here needs a rebuild — the changes are Python and shell.
One cell. This is the gate on everything after it.
bash scripts/run_paper003_capture_pilot.sh
Read CAPTURE VERDICT in the printed output.
| Verdict | Meaning | What to do |
|---|---|---|
CAPTURE |
Still, then rode, and the effect clears 20 mm | Go to stage 2 |
COLLISION |
The block moved but did not ride, or rode too little, or was never still | See the decision table below |
NONE |
The block never moved | The arm did not reach it — check ARM LAGGING first |
The verdict is computed with the same statistics the gate and arm D use
(carriage_evidence, estimate_capture), not a rule invented for the pilot. If
a trace is a capture by this measure it is a capture by the measure the arms are
scored through.
| Reason in the verdict | Likely cause | First thing to try |
|---|---|---|
| "still for only N steps" | The block is being nudged during the approach — the jaws are not straddling it | --grasp-radius smaller, or aim so the gap rather than a jaw meets it |
| "rode, but only X mm" | The gripper closed but lost the block, or the burst is too short | Check held_at_end in the JSON; raise --burst-on or --episode-steps |
| "no run of carriage steps" | The close did not take hold at all | Check grasp.closed_at is not null; try --gripper-close sign flipped |
ARM LAGGING present |
The arm cannot follow the script, so the geometry is wrong | Lower --script-speed or raise --approach-speed before anything else |
Do not raise the tolerance, widen the commit window, or lower
min_carriage_run to obtain a capture. Those are on the preregistration's
list of tuning that will not be done, and a capture obtained that way is not one.
nSEEDS=40 bash scripts/run_paper003_capture_pilot.sh
Read SUMMARY.txt:
Then read n off the table in the
preregistration, Sample size, and record it there
with this engagement figure beside it.
H3 needs the gate to fire on ≥ 0.90 of coupled trials and ≤ 0.10 of each control; H4 needs arm D not to regress on any of them.
for c in static noise drift slide; do
CONDITION=$c SEEDS=20 bash scripts/run_paper003_capture_pilot.sh
done
The one to watch is slide — a struck block that keeps its velocity is the
case that would collapse Paper 003 into Paper 002, and the gate leaked on it at
14–19% in the near-frictionless limit on CPU. Under real contact the friction is
whatever the scene's friction is, which is the point of measuring it here.
Also read gate_evidence per observation. A capture cell that fires through
proximity rather than carriage is not firing for the reason the design says
it should.
# from the laptop
bash scripts/copy_exp_surg_003_from_vessl.sh # adapt the paths for this run
Record under
experiments/surgical_intelligence/exp_surg_004_relation_expansion/results/
with a RESULTS.md that states, in this order: the capture verdict distribution,
engagement, normal_alignment, and the observation noise. Those are the four
things the preregistration is waiting on.
It is not the confirmatory sample and no cell from it may be pooled into
one. n is fixed from this pilot's engagement figure and the confirmatory cells
are run afterwards, on fresh seeds, against the frozen preregistration.
It is also not a test of the arms. If stage 1 says the scene produces a collision, the honest outcome is that the capture relation is not available in this scene — which is a result, and one the collision and carriage records already give the map for.