USB Gym: A Gymnasium Testbed for Closed-Loop Ultrasound View Matching

Thanapong Intharah; Zhichao Gao; Yiliang Liu; Hao Dong

cs.CVNov 5, 2026

USB Gym: A Gymnasium Testbed for Closed-Loop Ultrasound View Matching

Authors: Thanapong Intharah, Zhichao Gao, Yiliang Liu, Hao Dong

Organizations: Visual Intelligence Laboratory, Department of Statistics, Faculty of Science, Khon Kaen University, Khon Kaen 40002, Thailand · School of Computer Science, Peking University, Beijing 100871, China

Abstract

Background and Objective: Our main goal is ultrasound guidance: moving a probe until its image matches a target view. Published methods for each part of a guide are evaluated in separate simulators. We asked whether these methods, combined into one guide, are good enough. Methods: We built USB Gym, a testbed whose standard engine simulates ultrasound from public patient CT, with a network drawing organ textures from CT labels. Its guide has three replaceable parts (slots), an organ segmenter, a controller and a reward, which received 20 published methods, each run on the same 60 episodes of six held-out patients. In each slot, a published method replaced ours only if detectably better on these patients. The better of this combination and our original guide on 20 new CTs is the benchmark baseline. A pose oracle, reading the exact probe pose available only in simulation, set the reference. The results below are exploratory. Results: On these CTs, the probe started on average 16.0 mm from the target view; the pose oracle brought it to 1.5 mm. Of the results known when the combination was fixed, only an image-similarity reward (SSIM) was slightly but detectably better, so the combination was our guide with that reward. This combination was the benchmark baseline, ending at 13.7 mm (16% of the oracle's progress). Conclusions: Combining untuned published methods with ours did not guide the probe near the target view; the task remains under-explored. The testbed accepts new methods; no guide was tested on hardware.

Figures & tables

Environment	Image source	Closed loop	Components swapped (published methods)	Privileged state used	Robot or phantom test
Hase et al., 2020 [ 27 ]	Recorded frames of 34 volunteers on a grid	Yes, over the grid	Policy: the authors’ DQN variants	Grid position, in the reward	None reported
VesNet-RL, 2022 [ 28 ]	Simulated binary vessel images	Yes, in simulation	None described	Centerline distance, in the reward	Robot arm on vascular phantoms
SonoGym, 2025 [ 29 ]	Ray tracing or pix2pix on TotalSegmentator CT	Yes, in simulation	Policy (5); renderer (2)	Goal pose in the reward; true state for the expert	None reported
Isaac for Healthcare, 2025 [ 32 ]	OptiX ray tracing of labeled surface meshes	Yes; camera images or poses as input	Policy (2)	Target pose in a task’s observation and reward	None described
USB Gym (this work)	Physics render of CT label slices; optional texture stage	Yes, in simulation	Six slots (27; Table 2 )	Pose and CT masks, as rungs and in training; target-anchored contact †	None

Table 1: Runnable ultrasound navigation environments by publication year, as each source states them in October 2026. Privileged state: simulator state that a physical rig lacks, used by a reward, a teacher or the evaluation. † Privileged structure: the contact plane through the target window face and the start lift onto it (Section 3.3 ). Components swapped: parts swapped between runs, with the number of published methods run in them.

Panel of Figure 1 — Figure 1: The USB Gym closed loop. Each step slices the patient’s labeled CT at the probe pose, renders a frame and applies a texture stage. A mask source labels the organs of the observation and the target frame. The controller’s twist moves the probe, a lagged position servo, against a contact plane through the target window face. This target-anchored contact and the start lift onto it are privileged structure ( † , red dashed arrow). Shaded boxes are the six plug slots, and white boxes with a heavy outline are configured components (Table 2 ). The dashed privileged channel feeds evaluation code and imitation teachers only.

Term	Definition
Configurable component	A loop part that a user can replace; Table 2 lists eight.
Slot	A configurable component into which published methods were plugged; six of the eight.
Standard engine; guide	The engine is the physics renderer, the SonoSPADE texture stage and the reference CT labels; the guide is the mask source, controller and reward. All six slots are pluggable; Section 7 compares the guide’s.
Reference component	A component of the primary results, ours or a prior one, against which a slot’s plugged methods are compared.
Family	Pre-specified comparisons within which Holm’s adjustment runs; in a slot’s family, each tests a plugged method.
Secondary comparison	A reported comparison outside every family, with a $t$ CI and no decision.

Box 1 Terms and reading rules.

Component	Interface	Reference components	Plugged	Not evaluated, and why
Anatomy source	a labeled CT volume per patient	TotalSegmentator dataset masks [ 3 ] : 13 training and 6 held-out cases	1 method. Family ⋆ : TotalSegmentator v2 labels of the same CTs, fast mode	none
Acoustic windows	screening rules; target organs	screened windows; liver and kidney targets	none	none
Renderer	a label slice to a frame in $[0,1]$ ; physics-renderer parameters	physics renderer (authors’); DR off except in one ablation	2 methods. Family: LOTUS forward model, original code, published default parameters [ 12 ] ; the authors’ reimplementation of it. Secondary: SIMUS [ 10 , 11 ] , seed 0	Field II [ 8 ] : needs MATLAB. k-Wave [ 9 ] : needs an NVIDIA GPU, or hours per frame without one
Texture stage	a frame to a frame; a label-conditioned stage also reads the label slice	SonoSPADE [ 4 ] (authors’, method unchanged). Secondary: physics, CUT [ 36 ] , CycleGAN [ 14 ]	4 methods. Family: FastCUT [ 36 ] R , DCLGAN [ 37 ] R , UNSB [ 38 ] R , global histogram matching [ 39 ] R	Echo from Noise [ 40 ] , SDM [ 41 ] : no public abdominal weights. ConPres [ 42 ] : no public code
Mask source	a frame to an organ mask	U-Net [ 43 ] of the same stage and seed (authors’). Secondary: CT label †	7 methods. Family: LR-ASPP [ 44 ] , DeepLabV3 [ 45 ] on MobileNetV3-L or ResNet-50 [ 46 ] and FCN [ 47 ] , from scratch; nnU-Net [ 48 ] 2D, 2 epochs; SAM [ 49 ] and MedSAM [ 50 ] with U-Net boxes. Family ⋆ : SAMUS [ 51 ] with U-Net boxes; the torchvision networks with ImageNet-pretrained backbones. Secondary: SAM, MedSAM and SAMUS with CT-label boxes †	UltraSam [ 52 ] , USFM [ 53 ] : need an NVIDIA GPU
Contact	probe contact in MuJoCo, on or off	on; plane through the target window face, with the start lift onto it † ; press capped at 5 mm	none	none

Table 2: Configurable components of USB Gym: six plug slots, and the acoustic windows and contact, configured but not plugged. The anatomy source, renderer and texture stage form the standard simulation engine (results: Table 1 ); the mask source, controller and reward form the guide (Table 5 ). Interface: what a replacement provides. A slot’s pre-specified comparisons are made against its first reference component. Plugged: the published methods placed in the slot, led by their count, which takes each method once whatever its backbone, box prompt, pretrained weights or residual form, and leaves out ablations of

S

; the six counts sum to 27. Box 1 defines the entry types. † Privileged: a component that reads simulator state, or the target-anchored contact. ⋆ Specified after the first families were fixed, in families of its own. R The authors’ reimplementation of the published method. NCC: normalized cross-correlation. SSIM, NMI, LPIPS: structural similarity, normalized mutual information, learned perceptual image patch similarity. BC: behavior cloning. CEM: cross-entropy method.

Figure 2: The patient’s CT skin, its screened windows and the contact plane, shown for training patient s0011. (a) The skin surface extracted from the patient’s labeled CT, with the liver and kidneys inside. Dots mark the faces of the screened acoustic windows on the skin: 50 for the liver and 63 for the kidneys. Each stroke points along its window’s beam into the body. Hollow dots lie on the far side of the body from the viewer. (b) The probe seated at the top-ranked liver window, in the world frame of the MuJoCo scene. The contact plane, the top face of the static MuJoCo box, passes through the window face perpendicular to the beam axis. At this window the beam tilts $10^{\circ}$ from the skin normal toward the liver, so the plane is not tangent to the skin. Contact uses only this plane, not the skin surface. The fan is the slicing plane of the probe, painted with the CT labels it samples (orange liver, green gallbladder, blue vessel, grays other tissue). (c) The physics render at this pose and its CT label slice, mapped into the measured sector geometry without domain randomization and with a fixed speckle seed.

Machine	Apple M5 Pro, 24 GB, 15 CPU cores, macOS (Darwin arm64), Apple GPU through PyTorch MPS
Python, PyTorch	3.13.13, 2.12.1
NumPy, MuJoCo	2.1.3, 3.3.7
Gymnasium, Stable-Baselines3	1.0.0, 2.6.0

Table 3: Hardware and software of USB Gym’s runs: every reported result was computed on one Apple M5 Pro laptop with 24 GB of unified memory.

Final translation error (mm)
	Physics	CUT	CycleGAN	SonoSPADE L
Static start (no commanded motion)	16.03
Scripted sweep (no feedback)	15.52
Analytic pose oracle †	1.53
CT mask †
Servo base	11.65	11.65	11.65	11.65

Table 4: Oracle ladder: final translation error (mm), 6 held-out patients, contact on, evaluation domain randomization off; policy rows average 3 seeds, each meeting the same 60 crossed episodes. Pre-specified results, primary; the last block is post hoc. Errors: mean of per-patient means. Gaps: mean per-patient paired difference, patients of 6 sharing its sign, and 95% patient-cluster bootstrap CI (10,000 resamples); ∗ a difference under the pre-specified CI-and-sign rule. Every contrast is conditional on the trained artifacts (texture networks, segmenters, bases and policies). Seeds index artifacts; the patient is the sampling unit. The servo base accounts for nearly all of the improvement; imitation and REINFORCE each change the final translation error by less than 0.5 mm. L Reads the CT label slice. † Privileged: simulator state that no rig provides. Shaded: SonoSPADE, the texture stage of the standard engine and the main benchmark; physics, CUT and CycleGAN are the ladder’s other texture settings.

Slot and method	Error (mm)	Difference (mm), 95% $t$ CI, $k$ /6	Seeds	Holm $p$	TOST	Decision
Mask source . Condition: SonoSPADE frames; mask from the plugged source; servo base fitted under that mask, residual off; seeds 0 to 2, and seed 0 for the SAM-type rows, paired with seed 0. Reference: the stage-matched U-Net, 13.13 mm. Families: mask source ( $m=7$ ), second mask source ( $m=5$ ) and pretrained backbone ( $m=4$ ), never pooled. Post hoc MDD: 2.14 mm, 3.25 mm at $\alpha=0.05/7$ .
Reference
U-Net, stage-matched (authors’) [ 43 ]	13.13
Seed 0 only, for the seed-0 rows	13.33
Family ( $m=7$ )
LR-ASPP, MobileNetV3-L [ 44 ]	12.76	$-0.38$ $[-2.01,+1.25]$ , 4/6	2/3	$1.000$	no	no detectable difference

Table 5: Published methods in the three slots of the guide: mask source, controller and reward (exploratory, pre-specified rule). Every row: 12 steps, the 60 held-out episodes of Table 4 , contact on. Error: final translation error, mean over seeds of patient means. Difference: from the slot’s reference component, paired by seed, with 95% paired

t

CI and patients of 6 sharing its sign. Seeds: the seeds, of 3, whose own difference shares that sign (post hoc; n/a for seed-0 rows). Holm

p

: within the slot’s family. TOST: post hoc and unadjusted, yes when the 90% paired

t

CI lies inside

\pm 1

mm, the pre-specified equivalence margin. Decision: the pre-specified outcome words; e equivalence within

\pm 1

mm not shown. A decision of no detectable difference (in the pretrained-backbone family, does not detectably change) is not powered for equivalence; the TOST column shows where equivalence within

\pm 1

mm holds. a Unadjusted CI excludes 0; not declared a difference after Holm (post hoc note). Slot headers give the condition, the reference error and the post hoc minimum detectable difference (MDD) from the median SD of the slot’s contrasts with its reference, over the rows carried to the cohort analysis, so not SAM, MedSAM, CEM or the CT label, also at the first Holm level. The MDD is that of the paired

t

-test alone; the 5 of 6 sign condition makes the decision rule slightly less sensitive. Every contrast is conditional on the trained artifacts (texture networks, segmenters, bases and policies). Seeds index artifacts; the patient is the sampling unit. Costs,

t

and sign-flip

p

, the 90% CIs and scores are in the supplementary twin of this table. L Reads the CT label slice. † Privileged: simulator state that no rig provides. ‡ Fewer seeds than designed.

Slot and method	Error (mm)	Difference (mm), 95% $t$ CI, $k$ /6	Seeds	Holm $p$	TOST	Decision
Controller . Condition: SonoSPADE frames; CT mask † ; the plugged controller, residual or full twist as named; seeds 0 to 2, and seed 0 for CEM. Reference: the REINFORCE residual from the imitation start, 11.69 mm. Family: $m=10$ . Post hoc MDD: 2.09 mm, 3.41 mm at $\alpha=0.05/10$ .
Reference
REINFORCE residual, imitation start (authors’) [ 66 ]	11.69
Seed 0 only, for the seed-0 rows	11.79
Family ( $m=10$ )
PPO, residual [ 54 , 59 ]	11.50	$-0.19$ $[-0.53,+0.16]$ , 4/6	2/3	$0.885$	yes	no detectable difference

Table 6: Continued.

Analysis	Question	Outcome	Criterion met
Controls
Independent target speckle	Are appearance scores insensitive to shared speckle?	Largest: learned-mask servo base, physics, seed 0, 16.39 against 17.03 mm, $-0.63$ $[-2.04,+0.78]$ , 5/6; 4 conditions: $-0.63$ to $+0.05$ mm	yes
Positive control, SB3 learner	Can the RL training loop learn a correction?	2.40 to 2.58 mm in 3 of 3 seeds; bound 4.43, pose oracle 1.43, static start 15.45 mm	yes
Positive control, REINFORCE	Can REINFORCE remove a planted deficit?	Planted lateral deficit 3.84 mm; 0.32 to 0.38 of it removed, positive in 0 of 6 patients for every setting; the planted bias did not meet the reward-visibility rule	inconclusive
Label-channel controls	Does the SonoSPADE lead come from label access rather than texture?	Label-painted stage 14.22 mm against physics 17.20 and SonoSPADE 13.13 mm: $-2.97$ $[-5.50,-0.45]$ , 6/6 from physics, 0.73 [0.33, 1.34] of the lead; equivalence with SonoSPADE not shown. Rendered masks follow offset and permuted labels, 6/6 each. Permuted organ identities 21.69 mm, $+8.56$ $[+4.46,+12.65]$ , 6/6	inconclusive
Texture-budget test	Does the SonoSPADE lead persist with equal budgets and three texture seeds?	SonoSPADE, three texture seeds, 12.92 mm against budget-matched CUT 17.20, CycleGAN 16.93 and physics 17.20 mm: $-4.29$ $[-6.57,-2.00]$ , 6/6; $-4.01$ $[-6.74,-1.27]$ , 6/6; $-4.28$ $[-6.76,-1.80]$ , 6/6 mm	yes

Table 7: Controls, sensitivity analyses and replication: the question each answers and its outcome, absolute values first (exploratory). Supplementary Table S4.2 gives each criterion. Every contrast is conditional on the trained artifacts (texture networks, segmenters, bases and policies). Seeds index artifacts; the patient or CT is the sampling unit. Intervals: 95% paired

t

with the patients sharing the sign. Criterion met: yes or no, inconclusive, descriptive, or, for a family of changes, how many differ.

Panel of Figure 3 — Figure 3: Controls and replication: exploratory results, except the original six patients in (b), which are primary; daggers mark privileged inputs. (a) Top: SB3 PPO on the privileged easy variant, final error on training-patient episodes against environment steps, three seeds, with the static start, pass bound and pose oracle. Bottom: the REINFORCE control, the fraction of the planted lateral deficit left against iteration, three settings, pass line at one half. (b) Per-patient differences of the two cohort contrasts, original six and 20 cohort CTs, with means and 95% $t$ intervals, also for the original six, whose primary bootstrap intervals are in Table 4 . (c) Mean translation error against control step 0 to 36; the dotted line marks the evaluation horizon, step 12. Every contrast is conditional on the trained artifacts (texture networks, segmenters, bases and policies). Seeds index artifacts; the patient is the sampling unit.

Stage	Keep-best gain	Selection [CI]	Outcome	Directed control [CI]	Outcome
Vitale holdout [ 15 ] , liver, 12 subjects
SonoSPADE	moves the view	$+0.011$ [ $+0.002$ , $+0.024$ ]	not detectably better	$+0.041$ [ $-0.001$ , $+0.098$ ]	selection, not directed control
KidneyUS [ 73 ] , kidney, 55 subjects
SonoSPADE	moves the view	$+0.015$ [ $+0.003$ , $+0.028$ ]	informative selection signal	$+0.014$ [ $-0.011$ , $+0.037$ ]	selection, not directed control

Table 8: Outcomes of the real-frame cells of SonoSPADE, the standard engine’s texture stage (all eight testable cells, with the ladder’s other texture settings: Table 9 ). Keep-best gain: primary gain over the static start (Table 8 ). Selection: organ IoU of the state the match score selects along undirected motion, minus that of a random state. Directed control: the policy’s keep-best gain minus that of magnitude-matched undirected motion with the same selection. Values: exploratory per-subject IoU means [95% subject-by-CT bootstrap CI]. A cell holds if its CI excludes 0 and its Holm

p

over eight cells is below 0.05. In the holding cell, 38/55 subjects share the sign; exact sign-flip

p=3.4\times 10^{-4}

and Holm

p=0.003

. Elsewhere Holm

p

is 0.055 to 1.000.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: What the testbed renders: one held-out view under the four reference texture stages, from patient s0513 at the target pose of its first screened liver window (top) and kidney window (bottom). Left: the CT label map, with orange liver, green gallbladder, blue vessel, pink kidney and gray other tissue. Physics: the sector-warped render without domain randomization and with a fixed speckle seed, in which liver and surrounding tissue are nearly isoechoic (Section 3.4 ). CUT, CycleGAN and SonoSPADE act on that frame, and SonoSPADE also receives the label map. Supplementary Fig. S3.1 shows the same patient under the renderer-slot simulators. The figure is qualitative.

Row	Final (mm)	Keep-best (mm)
Reference ladder, 12 steps
Static start	16.03	16.03
Scripted sweep	15.52	15.57
Pose oracle †	1.53	1.49
CT-mask servo base † , four stages	11.65	11.25
CT mask † , after imitation, four stages	11.74	11.40

Appendix

Table 1: Keep-best error against final error, contact on (post hoc). Keep-best is the translation error at the step of highest match score

S

, a selection that reads no pose. The 12-step rows are the episodes of Table 4 , seeds pooled; four-stage rows average the four stages. The 36-step rows are exploratory. Plug rows use the episodes and seeds of Table 5 . † Privileged.

Figure 2: Mean translation error against control step on the six held-out patients, with contact on and no domain randomization at evaluation (post hoc). Lines show the static start, the pose oracle, the CT-mask servo base and final policies, and the learned-mask final policies. The CT-mask servos improve up to step 12, while three learned-mask policies are lowest at step 4 or 5 and then drift away. Supplementary Fig. S2.1 shows each patient’s final error on every rung. Every contrast is conditional on the trained artifacts (texture networks, segmenters, bases and policies). Seeds index artifacts; the patient is the sampling unit.

			Physics		CUT		CycleGAN		SonoSPADE
Knob	Default	Range	lo	hi	lo	hi	lo	hi	lo	hi
Overall gain	1	[0.8, 1.25]	0.44	0.38	0.38	0.32	0.48	0.43	0.21	0.18
Time-gain slope	1	[0.6, 1.4]	18.96	17.70	14.50	14.33	15.93	14.60	9.76	10.34
Lateral PSF $\sigma$ at focus (px)	1.1	[0.8, 1.6]	1.16	1.53	1.42	1.90	1.51	1.98	0.76	1.02
Focal depth (m)	0.05	[0.03, 0.08]	2.70	2.69	2.95	3.07	2.71	2.93	1.39	1.45
Speckle strength	0.9	[0.6, 1.2]	0.86	0.50	0.72	0.41	0.91	0.55	0.40	0.23

Appendix

Table 3: Domain-randomization (DR) knobs. When DR is on, each knob is drawn uniformly from its range once per episode and shared by the target and the observations. DR is off in training and evaluation for Tables 4 and 2 ; Table 4 reports the DR arms. Effect: mean absolute change of the in-sector pixel intensity (scale 0 to 1,

\times 100

) when the knob alone moves to the low (lo) or high (hi) end of its range, with the speckle seed fixed, over 60 frames (6 held-out patients

\times

10 jittered window poses), with each texture stage applied (physics: none). † Drawn independently per tissue label and measured one label at a time (mean over labels covering at least 1% of the sector). The speckle and noise seed is also redrawn per episode. The reference rows give the change from a new seed alone, a 5 mm in-plane slide and a full DR draw (5 draws per frame, 300 in all). The two weakest knobs, impedance multiplier and overall gain, change the frame by at most 0.21 and 0.48.

Rung	Stage	Mean	95% CI	$k$ /6
Contact off minus on
Static start	all	$0.00$	$[0.00,0.00]$	n/a
Scripted sweep	all	$+0.26$ ∗ H	$[+0.19,+0.33]$	6/6
Pose oracle (privileged)	all	$-0.17$ ∗	$[-0.25,-0.08]$	5/6
Servo base, CT mask	all	$+2.98$ ∗ H	$[+2.63,+3.29]$	6/6
Servo base, learned	Physics	$+0.88$ ∗	$[+0.22,+1.35]$	5/6

Appendix

Table 4: Contact and domain-randomization (DR) contrasts on the final pose error (mm), 6 held-out patients, 3 seeds, the crossed episodes of Table 4 . Contact rows: contact switched off at evaluation only, minus contact on, for every rung. The static start never moves, so its delta is 0 to numerical precision. The reference rungs and the CT-mask servo base do not depend on the texture stage and give one value (all). DR rows: the SonoSPADE CT-mask arm with DR in evaluation, in training, or in both, minus the same arm without DR (11.69 mm); these rows are descriptive. Each row gives the mean of per-patient paired differences, a 95% patient-cluster bootstrap CI and the number of patients that share the sign of the mean. ∗ : a difference under the CI-and-sign rule. H : also passes Holm on the paired

t

-test over the 16 contact rows (post hoc; Supplementary Table S2.18); 5 of the 14 starred rows do not.

		Translation (mm)			Rotation ( ∘ )
Rung	Contact	Lat.	Elev.	Axial	Out	In	Roll	Final	In contact	Gap	Lag
								(mm)	(share)	(mm)	(mm)
Start (all rungs)		11.05	9.54	3.64	9.44	8.30	1.89	16.03
Scripted sweep	on	11.03	9.71	1.10	3.57	3.10	1.89	15.52	0.88	$-0.78$	1.67
	off	11.01	9.68	2.71	9.67	8.30	1.89	15.78	0.00	$-2.75$	0.80
Pose oracle †	on	1.08	0.87	0.10	0.13	0.17	0.93	1.53	0.85	$-0.11$	4.60

Appendix

Table 5: Per-axis final error with contact on and off, 12 steps, the 60 held-out episodes (post hoc). Errors are patient-level means of the absolute SE(3)-log component in the target probe frame: lateral, elevational and axial translation, and out-of-plane tilt, in-plane tilt and roll. Base and policy rows pool seeds 0 to 2; four-stage rows also pool the four texture stages. In contact is the share of steps in contact. Gap is the mean height of the lowest face point above the plane, negative when pressed in. Lag is the mean distance between commanded and actual probe position after a step. The episodes are those of Table 4 . † Privileged.

Variant	W1 $\downarrow$	KID $\downarrow$	FID $\downarrow$	C2ST AUC
Baseline renderers
LOTUS, authors’ reimplementation	0.1272	$0.2288\pm 0.0066$	193.5	$0.982\pm 0.037$
NumPy renderer	0.1882	$0.3291\pm 0.0093$	256.8	$1.000\pm 0.002$
USB Gym stages, default render
Physics (no network)	0.0727	$0.2199\pm 0.0088$	198.7	$0.936\pm 0.102$
Physics + CUT	0.0277	$0.0740\pm 0.0045$	94.0	$0.701\pm 0.140$

Appendix

Table 6: Distance to real B-mode frames in matched sector geometry. Simulated side: 50 views per held-out patient (6 patients, 300 views) jittered around screened acoustic windows, warped to the real sector. Real side: the 60 annotated Vitale frames [ 15 ] , which no texture stage trained on. All scores use in-sector pixels only (sector mask eroded by 3 px). W1: intensity Wasserstein-1 distance. KID and FID: kernel and Fréchet Inception distances. C2ST AUC: area under the ROC curve of a CNN classifier two-sample test, mean

\pm

std over 5 repeats

\times

5 grouped folds (0.5 means the classifier cannot separate the domains). KID: mean

\pm

std over 100 random subsets of 50 frames per domain. Lower W1, KID and FID mean a smaller distance to the real frames. FID from 60 real frames is biased upward and is printed for completeness; KID is the unbiased kernel distance. NumPy renderer: the five-tissue appearance model of the authors’ earlier simulator (Supplementary Table S1.3). A post hoc label-permutation null with 200 permutations per stage gives a mean AUC of 0.50 (SD 0.07) for every stage and never reaches the observed AUC of any stage (

p=1/201

each). Physics W1 is below the authors’ reimplementation of LOTUS in 6/6 held-out patients and is below it pooled. In a post hoc comparison with the original LOTUS code, physics W1 is lower in 5/6 patients and lower pooled. Supplementary Table S2.16 gives the null per stage and post hoc frame-bootstrap intervals for W1 and KID. The established-simulator rows run the same protocol with the renderer swapped (exploratory). The original LOTUS code runs with its published default parameters.

Stage	Liver	GB	Vessel	Kidney	Spleen	Organs of 4
Real side: all 60 annotated Vitale frames, out-of-fold predictions
SonoSPADE	2/3	3/3	2/3	3/3	0/3	$3.3\pm 0.9$
Physics	3/3	3/3	3/3	1/3	0/3	$3.3\pm 0.5$
CUT	3/3	3/3	3/3	0/3	0/3	$3.0\pm 0.0$
CycleGAN	3/3	3/3	1/3	0/3	0/3	$2.3\pm 0.5$
Real side: the 18 Vitale holdout frames only

Appendix

Table 7: Cross-domain organ identity of each stage’s texture-matched segmenter. A cell counts the segmenter seeds, of 3, on which the modal predicted class of the organ is the correct class both in simulation (288 frames from the held-out patients) and on real frames. Liver, gallbladder (GB), vessel and spleen are scored on Vitale [ 15 ] : all 60 annotated frames, each predicted by a model that never trained on it (top), or only the 18 holdout frames that no segmenter trained on (bottom). Kidney is scored on the KidneyUS holdout [ 73 ] (107 frames) in both panels. Last column: organs of the four (liver, GB, vessel, kidney) per seed, mean

\pm

population std over the 3 seeds. Spleen is not a target and is descriptive only. The table is descriptive. Under the pre-specified rule, both panels detect no difference in organ-identity consistency across stages (not powered for equivalence).

Stage	Keep-best minus static IoU	95% CI	Rule met
Vitale holdout [ 15 ] , liver, $n=12$ targets
Physics	$+0.028$	$[+0.007,+0.051]$	yes
CUT	$+0.020$	$[+0.008,+0.033]$	yes
CycleGAN	$+0.030$	$[+0.014,+0.048]$	yes
SonoSPADE	$+0.062$	$[+0.029,+0.098]$	yes
Vitale holdout [ 15 ] , kidney, $n=4$ targets

Appendix

Table 8: Offline registration against real target frames (no probe, phantom or robot). Each real target is paired with 3 held-out CTs. Each learned-mask servo (3 seeds) takes its masks of the real frame and of the simulated view, for control and reward, from its own segmenter. Score: organ-restricted IoU between the CT label of the simulated view and the annotated organ of the real frame, at the keep-best step (chosen by that reward, without ground truth) minus at the static start. Mean over targets (each averaged over 9 runs), with a 95% bootstrap CI over targets. Rule met: yes only if the CI lies above 0 (pre-specified rule). AUL and KidneyUS frames are resized screenshots outside the convex-sector geometry of the renders.

Figure 3: Simulated and real frames in matched sector geometry. Each row pairs one annotated public real frame from the Vitale test set [ 80 , 15 ] with the held-out CT view whose organ layout best matches its annotation; the row label gives the frame, the CT patient and the organ-layout IoU of the match. For each real frame, the search scored 2,000 candidate probe poses per held-out patient around the screened acoustic windows by mean per-organ IoU and refined the best three by a cross-entropy search; a pose rotated by 180 degrees about the beam axis was also allowed. The search used label maps only, and no rendered image entered it. The columns show the real frame with its annotation outlined, the annotation, the matched CT label, and the physics, CUT, CycleGAN and SonoSPADE renders at the matched pose (no domain randomization, fixed speckle seed) with the CT label outlined (orange liver, green gallbladder, blue vessel, pink kidney, yellow spleen). The rows follow a fixed rule: the best frame of each organ stratum that has at least three frames, then the global IoU rank, with at most two rows per patient. Over all 60 frames, the median best IoU is 0.56 (interquartile range 0.42 to 0.81).

Stage	Keep-best gain	Selection [CI]	Outcome	Directed control [CI]	Outcome
Vitale holdout [ 15 ] , liver, 12 subjects
Physics	moves the view	$+0.008$ [ $-0.003$ , $+0.020$ ]	not detectably better	$+0.014$ [ $-0.023$ , $+0.052$ ]	selection, not directed control
CUT	moves the view	$+0.003$ [ $-0.007$ , $+0.014$ ]	not detectably better	$+0.010$ [ $-0.018$ , $+0.037$ ]	selection, not directed control
CycleGAN	moves the view	$-0.001$ [ $-0.012$ , $+0.010$ ]	not detectably better	$+0.023$ [ $-0.009$ , $+0.059$ ]	selection, not directed control
SonoSPADE	moves the view	$+0.011$ [ $+0.002$ , $+0.024$ ]	not detectably better	$+0.041$ [ $-0.001$ , $+0.098$ ]	selection, not directed control
KidneyUS [ 73 ] , kidney, 55 subjects

Appendix

Table 9: Outcomes of the eight testable real-frame cells, every texture stage of the ladder (Table 8 prints the rows of SonoSPADE, the standard engine’s stage). Keep-best gain: primary gain over the static start (Table 8 ). Selection: organ IoU of the state the match score selects along undirected motion, minus that of a random state. Directed control: the policy’s keep-best gain minus that of magnitude-matched undirected motion with the same selection. Values: exploratory per-subject IoU means [95% subject-by-CT bootstrap CI]. A cell holds if its CI excludes 0 and its Holm

p

over eight cells is below 0.05. In the holding cell, 38/55 subjects share the sign; exact sign-flip

p=3.4\times 10^{-4}

and Holm

p=0.003

. Elsewhere Holm

p

is 0.055 to 1.000.

Pipeline	Error (mm)	Minus static start (mm), 95% $t$ CI, $k$ /6	Minus pose oracle (mm), 95% $t$ CI, $k$ /6	20-CT cohort (mm)
Anchors (reference ladder, Table 4 )
Static start (no commanded motion)	16.03	n/a	$+14.50$ [ $+11.48$ , $+17.52$ ], 6/6	15.97
Scripted sweep (no feedback)	15.52	$-0.51$ [ $-0.81$ , $-0.22$ ], 5/6	$+13.99$ [ $+11.12$ , $+16.86$ ], 6/6	15.38
Analytic pose oracle †	1.53	$-14.50$ [ $-17.52$ , $-11.48$ ], 6/6	n/a	1.47
No privileged input at run time: frames without a label-conditioned stage, learned mask
Reference: Physics; U-Net; servo base	17.20	$+1.17$ [ $-1.21$ , $+3.54$ ], 4/6	$+15.67$ [ $+13.94$ , $+17.39$ ], 6/6	15.75

Appendix

Table 10: Lowest measured error per kind of run-time input, post hoc. Post hoc: lowest point estimate on the same six held-out patients that measured it, so optimistic; not the pre-specified combination rule. The pre-specified combination selects by the slot decisions and tests its choice on 20 CTs that played no part in it, so the two can name different pipelines. Pipelines name the frames, the mask source and the controller. Error: final translation error over the 60 held-out episodes of Table 4 , mean over seeds of patient means. Differences are paired by patient, with the 95%

t

CI and the patients of 6 sharing the sign. A random-action controller, from earlier runs with no pre-specified rule, ends at 17.45 mm (3 seeds; Supplementary Table S3.2). Each post hoc lowest row is the lowest error among the 18, 17 and 43 measured candidates of the second, third and fourth blocks. No candidate combines the winners of two slots; the pre-specified combination is reported in Section 7.4 . The learned-mask controller arm with new learner seeds is not a candidate; its rows are in Table 2 . Both lowest rows without privileged input are BC residuals (exploratory, no pre-specified rule). None of these rows enters a decision, and none is read as a separation of two methods. Cohort: the same configuration on the 20-CT replication cohort, an unselected check; n/r, not run there. Every episode keeps the contact plane through the target window face and a start drawn around the target. No row is therefore free of target information. L Label-conditioned stage: its network also reads the CT label slice. † Privileged: simulator state that no rig provides. E Exploratory, no pre-specified rule.

Figure 1: Published methods in the plug slots, each drawn as its difference from its slot’s reference component, the solid vertical line. Each line is one plugged method of the five slot families of Table 5 . The thin bar is the 95% paired

t

CI, and the thick bar is the 90% paired

t

CI. A filled marker means that the method is equivalent to its reference within

\pm 1

mm, the pre-specified margin, which the dotted lines mark (post hoc, unadjusted TOST). The right column prints each difference with its 95% CI and outcome word, and daggers mark slots under the privileged CT mask. These results are exploratory; Table 5 gives each slot’s post hoc minimum detectable difference. Every contrast is conditional on the trained artifacts (texture networks, segmenters, bases and policies). Seeds index artifacts; the patient is the sampling unit.

Slot and method	Error (mm)	Difference (mm), 95% $t$ CI, $k$ /6	Seeds	Holm $p$	TOST	Decision
Texture stage . Condition: frames from the plugged stage; mask from that stage’s own matched U-Net; servo base, residual off; seeds 0 to 2. Reference: SonoSPADE, 13.13 mm. Family: $m=4$ . Post hoc MDD: 3.24 mm, 4.39 mm at $\alpha=0.05/4$ .
Reference
SonoSPADE (authors’) [ 4 ] L	13.13
Family ( $m=4$ )
FastCUT [ 36 ]	16.13	$+3.00$ $[+0.24,+5.75]$ , 5/6	3/3	$0.038$	no	higher
DCLGAN [ 37 ]	16.94	$+3.80$ $[+1.36,+6.25]$ , 5/6	3/3	$0.021$	no	higher

Appendix

Table 1: Published methods in the three slots of the standard engine: texture stage, renderer and CT anatomy (exploratory, pre-specified rule). The engine is the physics renderer with the SonoSPADE texture stage and the reference CT labels; every slot stays pluggable. Columns, slot headers, marks and notes as in Table 5 ; the MDD leaves out SIMUS and the painted stage.

Contrast	Family	Difference or 90% CI (mm)	$k$ /6	Test $p$	Sign-flip $p$	Holm $p$	Decision	Crossed bootstrap 95% CI
B1: DAgger residual minus learned-mask servo base, physics frames; the DAgger residual ends at 14.02 mm	LM-D	$-3.17$ $[-5.04,-1.31]$	6/6	$0.007$ (paired $t$ )	$0.031$	$0.043$	holds lower	$[-4.56,-1.26]$
Equivalence within $\pm 1$ mm	LM-E	90% $[-4.63,-1.71]$		$0.985$ (TOST)		$1.000$	equivalence not shown
B2: DAgger residual minus learned-mask servo base, SonoSPADE frames; the DAgger residual ends at 12.04 mm	LM-D	$-1.09$ $[-2.70,+0.51]$	5/6	$0.141$ (paired $t$ )	$0.156$	$0.564$	no detectable difference	$[-2.96,+0.28]$
Equivalence within $\pm 1$ mm	LM-E	90% $[-2.35,+0.17]$		$0.555$ (TOST)		$1.000$	equivalence not shown
B3: BC residual minus learned-mask servo base, physics frames; the BC residual ends at 14.29 mm	LM-D	$-2.91$ $[-5.52,-0.30]$	6/6	$0.035$ (paired $t$ )	$0.031$	$0.177$	no detectable difference	$[-5.45,-0.55]$
Equivalence within $\pm 1$ mm	LM-E	90% $[-4.95,-0.86]$		$0.940$ (TOST)		$1.000$	equivalence not shown

Appendix

Table 2: Learned-mask controller experiment on the six held-out patients (exploratory, pre-specified rule). Families LM-D (difference) and LM-E (equivalence within

\pm 1

mm),

m=6

each, Holm within each family. A difference holds if its Holm

p

is below 0.05 and at least 5 of 6 patients share its sign (paired

t

, df 5); the exact sign-flip

p

is printed and never decides. Absolute errors are the base’s mean plus the paired difference. The servo base ends at 17.20 mm with physics frames and 13.13 mm with SonoSPADE frames, and the pose oracle at 1.53 mm. MDD from the observed SDs: 2.37 and 3.63 mm for the base contrasts, 2.69 and 4.12 mm for the learner pair, at

\alpha=0.05

and at Holm’s first step. The separation criterion (a physics-frame difference of B1, B3 or P1 holds) is met. Crossed bootstrap: case by seed, a sensitivity check.

Rung	Family	Change or 90% CI (mm)	$k$ /6	Test $p$	Sign-flip $p$	Holm $p$	Decision	Crossed bootstrap 95% CI
Static start	SKIN-D	$+0.36$ $[-0.55,+1.28]$	2/6	$0.353$ (paired $t$ )	$0.500$	$0.620$	no detectable change	n/a
Equivalence within $\pm 1$ mm	SKIN-E	90% $[-0.35,+1.08]$		$0.067$ (TOST)		$0.337$	equivalence not shown
Scripted sweep	SKIN-D	$+0.55$ $[-0.70,+1.80]$	4/6	$0.310$ (paired $t$ )	$0.250$	$0.620$	no detectable change	n/a
Equivalence within $\pm 1$ mm	SKIN-E	90% $[-0.43,+1.53]$		$0.198$ (TOST)		$0.792$	equivalence not shown
Pose oracle (privileged)	SKIN-D	$+0.96$ $[-0.20,+2.12]$	6/6	$0.086$ (paired $t$ )	$0.031$	$0.257$	no detectable change	n/a
Equivalence within $\pm 1$ mm	SKIN-E	90% $[+0.06,+1.87]$		$0.470$ (TOST)		$1.000$	equivalence not shown

Appendix

Table 3: Skin-tangent contact experiment (exploratory, pre-specified rule): skin-tangent plane minus target-anchored plane, same rung and seed, on the six held-out patients. Families SKIN-D (change) and SKIN-E (equivalence within

\pm 1

mm),

m=5

each, Holm within each family; a change holds if its Holm

p

is below 0.05 and at least 5 of 6 patients share its sign (paired

t

, df 5); the exact sign-flip

p

is printed and never decides. Both servo bases keep the calibrations fit with the target-anchored plane. MDD from the observed SDs: 1.71 mm at

\alpha=0.05

and 2.51 mm at Holm’s first step. Lower block: the plane quantities over the same sampled setups, per rung (n episodes: 60 setups, times 3 seeds for the bases).

Split, slice spacing	Screened	Pass	Rejected	Label check failed	Liver windows, median	Kidney windows, median	No kidney window
Tr, slice 2.5 mm	4	4	0	0	10.5	30.0	0
Tr, slice 2.0 mm	34	32	2	1	19.0	35.0	1
Va, slice 2.0 mm	33	33	0	0	22.0	38.0	0
Tr, slice 1.25 mm	3	2	1	1	17.5	36.0	0
Va, slice 2.5 mm	3	3	0	0	21.0	36.0	0
Va, slice 1.25 mm	2	2	0	0	22.0	39.0	0

Appendix

Table 4: Yield of the AMOS external cohort (exploratory). Of 79 eligible AMOS CTs (slice spacing at most 2.5 mm), 77 passed the label check and 76 the window screen. The first 30 passes in the seeded order form the cohort. Rows: AMOS split (Tr, training; Va, validation) and native slice spacing. Windows: screened acoustic windows per CT. Every AMOS structure comes from TotalSegmentator v2 fast-mode predictions; the AMOS manual labels serve only to accept a case, which needs a liver Dice of at least 0.85 and a kidney Dice of at least 0.70, with 5 mL as the absent-kidney volume.

Structure	Accepted: median (min, max)			Rejected: median (min, max)
Spleen	0.94	(0.91,	0.96)	0.95	(0.94,	0.95)
Right kidney	0.92	(0.00,	0.95)	0.72	(0.53,	0.95)
Left kidney	0.93	(0.87,	0.95)	0.80	(0.55,	0.93)
Gallbladder	0.82	(0.00,	0.93)	0.84	(0.70,	0.88)
Esophagus	0.80	(0.50,	0.87)	0.81	(0.67,	0.82)
Liver	0.96	(0.88,	0.97)	0.94	(0.90,	0.97)

Appendix

Table 5: Agreement of the TotalSegmentator v2 labels with the AMOS manual organ labels, Dice per structure (descriptive). Accepted: the 30 cohort CTs; rejected: the 3 screened CTs rejected by the label check or the window screen. Median over the accepted CTs: liver 0.96, kidneys 0.93.

Rung	Original six, TotalSegmentator labels (mm)	AMOS (mm), 95% $t$ CI	AMOS $n$
CT-mask servo base, physics	11.73	12.18 $[11.46,12.89]$	30
Learned-mask servo base, physics	16.58	16.63 $[16.02,17.24]$	30
Learned-mask servo base, SonoSPADE	12.62	13.91 $[13.16,14.65]$	30
CT-mask policy, physics, final weights	11.70	12.36 $[11.65,13.07]$	30
CT-mask policy, SonoSPADE, final weights	11.62	12.24 $[11.52,12.96]$	30

Appendix

Table 6: CT-anatomy context (descriptive): the five CT-anatomy rungs of the original six on TotalSegmentator labels beside the same rungs on the 30 AMOS CTs. The two columns share the label model and differ in CT source, patients,

n

, the labels of the calibration fit and the window source; they are never pooled, and no difference across them is printed. Error: final translation error, mean over seeds of case means.

Contrast	Estimate (mm), 95% $t$ CI	$k$ / $n$	Sign-flip $p$	Holm $p$	Decision	Case bootstrap	Case-by-seed bootstrap	Six
C1: learned minus CT masks, physics, final policy	$+4.01$ $[+3.40,+4.62]$	29/30	$1.0\times 10^{-6}$	$3.0\times 10^{-6}$	replicates	$[+3.42,+4.56]$	$[+3.29,+4.68]$	$+4.92$
C2: SonoSPADE minus physics, learned-mask servo base	$-2.73$ $[-3.43,-2.03]$	27/30	$1.0\times 10^{-6}$	$3.0\times 10^{-6}$	replicates	$[-3.39,-2.06]$	$[-3.50,-1.97]$	$-4.06$
Gap: CT-mask servo base minus pose oracle	$+10.68$ $[+9.99,+11.37]$	30/30	$1.0\times 10^{-6}$	$2.0\times 10^{-6}$	holds higher	$[+10.05,+11.33]$	$[+9.99,+11.40]$	$+10.12$
Bound: privileged look-ahead (seed 0) minus CT-mask servo base	$-4.51$ $[-5.29,-3.74]$	30/30	$1.0\times 10^{-6}$	$2.0\times 10^{-6}$	holds lower	$[-5.24,-3.80]$	$[-5.31,-3.74]$	$-3.52$
C3: RL change within $\pm 1$ mm in all 8 policy arms (intersection-union TOST)			$0.009$	$0.009$	equivalence within $\pm 1$ mm shown in every arm
Largest per-arm TOST $p$ : learned masks, CUT	$+0.52$ , 90% $[+0.20,+0.84]$	20/30	$0.009$ (TOST)

Appendix

Table 7: AMOS external cohort on 30 CTs (exploratory, pre-specified rule; external replication in simulation only). Top: the decisions. Families AMOS-P (

m=3

) and AMOS-S (

m=2

), two-sided Monte Carlo sign-flip test (1,000,000 sign vectors, seed 0), Holm within each family; the CT is the unit. C1 and C2 replicate when they hold with the sign of the original six. Middle: AMOS beside the 20 replication-cohort CTs and the original six (estimation only; no

p

, no decision; 95%

t

CIs). An interval that covers 0 does not show that a value is unchanged, and one that excludes 0 is not a difference that holds. Bottom: the relabeling control on the 20 cohort CTs (estimation only), the paired change and the AMOS difference side by side, never added.

Contrast (relabeling control on 20 cohort CTs)	Relabeled minus cohort, paired 95% $t$ CI	$k$ / $n$	AMOS minus relabeled cohort, Welch 95% CI
C1: learned minus CT masks, physics, final policy	$+0.84$ $[-0.03,+1.70]$	13/20	$+0.61$ $[-0.71,+1.92]$
C2: SonoSPADE minus physics, learned-mask servo base	$-0.61$ $[-1.43,+0.20]$	11/20	$-0.07$ $[-1.17,+1.03]$
Gap: CT-mask servo base minus pose oracle	$-0.54$ $[-1.21,+0.12]$	15/20	$+0.05$ $[-1.10,+1.20]$
Bound: privileged look-ahead (seed 0) minus CT-mask servo base	$-0.38$ $[-1.39,+0.63]$	11/20	$+0.82$ $[-0.69,+2.34]$
RL change, CT masks, physics	$+0.04$ $[-0.12,+0.19]$	11/20	$-0.01$ $[-0.19,+0.16]$
RL change, CT masks, CUT	$+0.04$ $[-0.08,+0.16]$	10/20	$-0.05$ $[-0.22,+0.11]$

Appendix

Table 8: Continued.

Contrast	Family	Difference or 90% CI (mm)	$k$ /20	Sign-flip $p$	Holm $p$	Decision	Crossed 95% CI	MDD (mm)
Pipelines on the 20 cohort CTs, final translation error (mm), as measured and unranked
Static start 15.97; pose oracle (privileged) 1.47
With physics frames: authors’ original pipeline 15.36; its learned-mask servo base 15.75; the combination 15.62
With SonoSPADE frames: authors’ original pipeline 13.96; its learned-mask servo base 13.71; the combination 13.70
Scripted sweep 15.38; CUT frames: learned-mask servo base 16.31, REINFORCE final 15.99; CycleGAN frames: base 16.14, final 15.89
Families on the cohort CTs

Appendix

Table 9: Combination chosen by the pre-specified rule (exploratory): the authors’ original pipeline with the SSIM appearance term in the REINFORCE reward, the only slot the rule changed. Primary: the 20 replication-cohort CTs, exact sign-flip test, Holm within families CMB-D (

m=4

, difference), CMB-E (

m=4

, equivalence within

\pm 1

mm, TOST) and CMB-G (

m=2

, gap to the pose oracle, conditional on this renderer: grid-fixed speckle, no elevational beam width, rigid tissue). Each D difference compares three trained policies per arm and is conditional on those six RL runs. Other pipelines measured on these CTs print as measured and none is ranked against the combination. MDD from the observed SDs, at

\alpha=0.05

and at Holm’s first step. The continuation prints the six held-out patients (selection-biased; estimation only), the same-procedure retraining of the match score

S

(estimation only) and AMOS (secondary; estimation only).

Row	Estimate (mm), 95% $t$ CI	$k$ / $n$	Sign-flip $p$	Interval (sensitivity)
Secondary, cohort CTs (no family, no decision)
Combination, physics frames: mean final error	15.62 $[14.76,16.48]$
minus the static start	$-0.35$ $[-1.20,+0.49]$	13/20
share of the static-to-oracle distance closed	0.024 $[-0.031,0.076]$ (case bootstrap)
Combination, SonoSPADE frames: mean final error	13.70 $[12.68,14.73]$
minus the static start	$-2.27$ $[-3.31,-1.22]$	18/20

Appendix

Table 10: Continued.

Slot and row	Difference (mm), 95% $t$ CI, $k$ /6	Per seed (mm)	Seeds	Patient-by-seed 95% CI	Crossed 95% CI
Texture stage , minus SonoSPADE
None, physics frame (authors’)	$+4.06$ $[+1.68,+6.44]$ , 5/6	$+3.70,+3.89,+4.60$	3/3	$[+2.09,+5.70]$	$[+1.99,+5.75]$
CUT (prior) [ 36 ]	$+3.70$ $[+1.33,+6.07]$ , 5/6	$+3.44,+4.41,+3.26$	3/3	$[+1.51,+5.73]$	$[+1.37,+5.72]$
CycleGAN (prior) [ 14 ]	$+3.72$ $[+1.18,+6.26]$ , 5/6	$+3.28,+4.52,+3.35$	3/3	$[+1.44,+5.60]$	$[+1.35,+5.68]$
Label-painted physics stage (authors’, not learned)	$+1.09$ $[-1.98,+4.16]$ , 3/6	$+1.66,+2.11,-0.49$	2/3	$[-1.27,+3.46]$	$[-1.53,+3.80]$
FastCUT [ 36 ]	$+3.00$ $[+0.24,+5.75]$ , 5/6	$+3.23,+3.68,+2.08$	3/3	$[+0.81,+5.00]$	$[+0.72,+5.15]$

Appendix

Table 11: Patient-by-seed intervals of the Table 5 rows with three seeds (exploratory). Difference and

k

/6 as in Table 5 . Per seed: the difference of patient means within one seed, seed

s

of the row minus seed

s

of the reference, for seeds 0, 1 and 2. Seeds: those whose difference shares the sign of the mean (post hoc). Patient-by-seed: the pre-specified bootstrap, which resamples patients, then seeds within each sampled patient, with the seed index paired across arms; 10,000 resamples, seed 0. Crossed: patients and seeds resampled independently, a post hoc sensitivity check with the same resamples and seed. Rows run at seed 0 only have no seed factor and are not listed. Every contrast is conditional on the trained artifacts (texture networks, segmenters, bases and policies). † Privileged.

Slot and row	Difference (mm), 95% $t$ CI, $k$ /6	Per seed (mm)	Seeds	Patient-by-seed 95% CI	Crossed 95% CI
Controller , minus the REINFORCE residual from the imitation start
Servo base alone (authors’)	$-0.04$ $[-0.36,+0.29]$ , 3/6	$-0.34,0.00,+0.23$	2/3	$[-0.31,+0.25]$	$[-0.44,+0.29]$
REINFORCE residual, BatchNorm in eval mode (authors’)	$+0.01$ $[-0.28,+0.30]$ , 3/6	$-0.01,-0.07,+0.12$	1/3	$[-0.22,+0.28]$	$[-0.24,+0.29]$
PPO, residual [ 54 , 59 ]	$-0.19$ $[-0.53,+0.16]$ , 4/6	$-0.58,+0.20,-0.18$	2/3	$[-0.50,+0.17]$	$[-0.59,+0.33]$
SAC, residual [ 55 , 59 ]	$+0.37$ $[-0.86,+1.61]$ , 4/6	$-0.29,+1.94,-0.53$	1/3	$[-0.92,+1.58]$	$[-1.36,+2.02]$
TD3, residual [ 56 , 59 ]	$+8.54$ $[+7.00,+10.08]$ , 6/6	$+6.30,+8.41,+10.90$	3/3	$[+6.60,+10.09]$	$[+5.63,+10.80]$

Appendix

Table 12: Continued.

Slot and method	Role	Error (mm)	Difference (mm), 95% $t$ CI, $k$ /6	Decision
Texture stage . Condition: frames from the plugged stage; mask from that stage’s own matched U-Net; servo base, residual off; seeds 0 to 2. Reference: SonoSPADE, 13.13 mm.
SonoSPADE (authors’) [ 4 ] L	reference	13.13
Label-painted physics stage (authors’, not learned) L	secondary	14.22	$+1.09$ $[-1.98,+4.16]$ , 3/6	secondary
FastCUT [ 36 ]	family	16.13	$+3.00$ $[+0.24,+5.75]$ , 5/6	higher
CUT (prior) [ 36 ]	secondary	16.84	$+3.70$ $[+1.33,+6.07]$ , 5/6	secondary
CycleGAN (prior) [ 14 ]	secondary	16.85	$+3.72$ $[+1.18,+6.26]$ , 5/6	secondary

Appendix

Table 1: Leaderboard of the methods plugged into each slot (exploratory). Each row is the row of Table 5 (the guide’s slots) or Table 1 (the engine’s slots) for the same method, run in its slot’s condition (slot headers, as in those tables), with 12 steps, the 60 held-out episodes of Table 4 and contact on. In each slot the reference comes first, then the rows whose difference is taken from it, by mean error, lowest first, with the privileged rows last. Rows run at seed 0 only follow the reference’s seed-0 line in the same order, since their differences are taken from that line and their error is that of one seed. The order is by mean error only and is not a ranking: each row was tested against its slot’s reference, not against the other rows, and where the decision is no detectable difference, the methods are not distinguished from the reference or from each other. The CT-anatomy rows keep the order of Table 1 , since each is scored against its own reference level. Authors’: our published component [ 5 ] . Role: the row’s group in those tables; second family: the second mask-source family or the LPIPS reward family; privileged: its group of privileged rows, which are secondary. Error, difference with its 95% paired

t

CI and

k

/6, and decision, with the marks L,

\dagger

\ddagger

, a and e, as in those tables. The pretrained-backbone differences of Table 5 , which have no error of their own, are left out.

Slot and method	Role	Error (mm)	Difference (mm), 95% $t$ CI, $k$ /6	Decision
Controller . Condition: SonoSPADE frames; CT mask † ; the plugged controller, residual or full twist as named; seeds 0 to 2, and seed 0 for CEM. Reference: the REINFORCE residual from the imitation start, 11.69 mm.
REINFORCE residual, imitation start (authors’) [ 66 ]	reference	11.69
DAgger, residual [ 61 ]	family	10.85	$-0.84$ $[-3.42,+1.74]$ , 4/6	no detectable difference
BC, residual [ 60 ]	family	11.25	$-0.44$ $[-1.42,+0.55]$ , 4/6	no detectable difference
PPO, residual [ 54 , 59 ]	family	11.50	$-0.19$ $[-0.53,+0.16]$ , 4/6	no detectable difference
Servo base alone (authors’)	secondary	11.65	$-0.04$ $[-0.36,+0.29]$ , 3/6	secondary

Appendix

Table 2: Continued.

Robotic Ultrasound

Robotic ultrasound uses robots to acquire and interpret ultrasound images for autonomous scanning and image-guided procedures.

See all 28 papers

Action-Conditioned World Model for Goal Plane Guidance in Robotic Neck Ultrasound
Jul 24, 2026Siqi Fan, Ran Liu, Zixuan Yang +6World Model Learning Action-Conditioned World Models
Department of Mechanical and Automation Engineering, Chinese University of Hong Kong, Hong Kong SAR · Centre for Artificial Intelligence and Robotics Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, Hong Kong SAR · Department of Vascular Ultrasound, Xuanwu Hospital, the Capital Medical University +3
SonoSelect: Efficient Ultrasound Perception via Active Probe Exploration
Apr 7, 2026Yixin Zhang, Yunzhong Hou, Longqi Li +3Active View Selection Active Sensing
Shandong University, China · Beijing Institute of Technology, China · Yale University, USA +1
Ultrasound Operator Guidance Using World Modeling and Retrieval Based Action Planning
Oct 5, 2026Noortje I. P. Schueler, Hans van Gorp, Ruud J. G. van SlounWorld Model-Based Planning Ultrasound Imaging
Department of Electrical Engineering, Eindhoven University of Technology, 5612 AZ Eindhoven, The Netherlands

Jul 24, 2026J/K move · Enter open · S save

Action-Conditioned World Model for Goal Plane Guidance in Robotic Neck Ultrasound

Siqi Fan, Ran Liu, Zixuan Yang, Xiaoyu Fu+5

Department of Mechanical and Automation Engineering, Chinese University of Hong Kong, Hong Kong SAR · Centre for Artificial Intelligence and Robotics Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, Hong Kong SAR · Department of Vascular Ultrasound, Xuanwu Hospital, the Capital Medical University+3

World Model Learning Action-Conditioned World Models Robotics Robotic Ultrasound Goal-Conditioned RL

Open paper AskPDF arXiv

Ultrasound examinations typically follow specific scanning protocols that require navigating through an ordered sequence of anatomical landmarks and target imaging planes. We present a local action-conditioned world model framework for goal plane probe guidance in robotic ultrasound. Autonomous ultrasound tasks often require large numbers of probe-motion trajectories for training, but collecting high-quality demonstrations is labor-intensive and explicit simulators are difficult to build because ultrasound appearance depends on contact, tissue deformation, and view-dependent acoustic artifacts. We address this problem with a two-stage model-based learning pipeline. First, a latent conditional diffusion world model predicts future ultrasound observations from recent context frames, probe motions and temporal offset. Second, a goal-conditioned temporal transformer predicts ordered probe motions and is fine-tuned using rewards from the frozen world model. Experiments on the self-collected dataset show that the world model preserves action-dependent anatomical structure on target-directed scans. In real-world closed loop experiments, the framework achieves success rates of 70.0% for carotid guidance and 65.0% for thyroid guidance. These results demonstrate the potential of learned ultrasound dynamics for training goal-directed robotic probe navigation within protocol-constrained organ specific tasks.

Abstract

Robotic Ultrasound

Explore similar work

Action-Conditioned World Model for Goal Plane Guidance in Robotic Neck Ultrasound

SonoSelect: Efficient Ultrasound Perception via Active Probe Exploration

Ultrasound Operator Guidance Using World Modeling and Retrieval Based Action Planning