UltraDP: Diffusion Control in Carotid Ultrasound

• UltraDP is a multimodal, force-aware diffusion-policy system that autonomously performs transverse carotid ultrasound scanning with a 95% success rate.
• It uses a denoising diffusion model with an inference-time guidance mechanism to generate multimodal actions that maintain artery centering during scanning.
• The system leverages expert demonstrations, ultrasound landmark pretraining, and hybrid force-impedance control to ensure safe, robust, and efficient scanning outperforming baseline methods.

UltraDP is a force-aware diffusion-policy method for autonomous transverse carotid ultrasound scanning. It receives multimodal observations—ultrasound images, wrist camera RGB-D images, contact wrench, and probe pose—and generates actions suited to the multi-modal action distributions that arise in robotic ultrasound, while a hybrid force-impedance controller realizes those actions during physical interaction with a human subject. The system combines expert-demonstration learning, a pretrained ultrasound landmark encoder, an inference-time guidance mechanism for artery centering, and a large real-world dataset of carotid scans; on previously unseen subjects it achieves a 95% success rate in transverse scanning (Chen et al., 19 Nov 2025).

1. Clinical task and problem formulation

UltraDP is designed for autonomous transverse-section carotid scanning. The scan begins near the lower neck and proceeds upward while the robot maintains acoustic contact, produces diagnostically useful transverse ultrasound images of the carotid artery, and terminates when the carotid bifurcation is observed. A central task requirement is that the artery remain horizontally centered in the ultrasound image, because this facilitates subsequent longitudinal scanning. The operational objective therefore includes maintaining contact and image quality, moving upward along the neck, keeping the carotid artery centered horizontally in the image, recognizing the bifurcation, and doing so safely and comfortably on unseen human subjects (Chen et al., 19 Nov 2025).

The paper presents this setting as a difficult human-in-the-loop robotic perception-and-control problem. Neck anatomy varies across subjects in body habitus, neck length, artery angle, offsets, and tissue compliance. Ultrasound imaging is itself contact-sensitive: image quality depends on pressure, orientation, tilt, and coupling, so similar observations can admit multiple valid corrective actions. The environment is a living human rather than a rigid scene, and both external wrench and feasible motion range vary online. The authors further emphasize that successful control requires a kind of sonographer-like “spatial imagination,” namely the ability to map 2D slices to 3D anatomy and choose corrective actions accordingly.

Within this framing, UltraDP is positioned against two prior classes of systems. Rule-based robotic ultrasound methods rely on engineered perception and hand-tuned heuristics such as image centering or finite-state logic, but are brittle when landmark detection fails or calibration assumptions do not transfer. Prior learning-based methods, including behavior cloning and world-model-style approaches, can use demonstrations but are described as limited in generalization and especially in handling the multi-modal action distributions of ultrasound scanning. This is the principal motivation for a diffusion-policy formulation.

2. Dataset, demonstrations, and landmark pretraining

The training data are collected from expert demonstrations in which a sonographer physically holds the manipulator flange while the system records synchronized multimodal observations. At time $t$, the observation is

$\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$

where the ultrasound image is

$\bm U \in \mathbb{R}^{H_1 \times W_1},$

the wrist RGB-D image is

$\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$

the probe pose is

$\bm x \in SE(3),$

and the contact wrench is

$\bm w \in \mathbb{R}^6.$

A complete demonstration trajectory is written as

$$\mathcal{D}=\{d_1,d_2,\cdots,d_n \mid d_i = (\bm U_t, \bm I_t,\bm x_t,\bm w_t)_{t=1:T_n}\},$$

and each time step is paired with the action

$$\bm A_t = \left(diff(\bm x_{t+1},\bm x_t),\bm w_t\right),$$

so the action contains both a relative pose increment and the desired contact wrench (Chen et al., 19 Nov 2025).

The main dataset contains 21 volunteers, 210 scans, and 460k sample pairs, with volunteers of both genders and demonstrations collected by two certified sonographers. Multiple trajectories were recorded per volunteer with slight variations in position and posture. For additional validation of generalization, the authors collected 54k sample pairs from new male and female volunteers not seen in training. Data augmentation is performed by diffusing each demonstration trajectory with random transformation noise in Cartesian space; for wrist RGB-D images, a depth-based mask removes irrelevant background. The end-effector pose is represented by a 3D position plus a 6D rotation representation following Zhou et al. The paper does not provide an explicit numeric train/validation/test split, but its evaluation protocol separates training subjects from previously unseen volunteers.

A distinct pretraining stage is applied to the ultrasound encoder. A ResNet is trained on a separate dataset of 6k ultrasound image-label pairs $(\bm U, l)$, where $l \in \mathbb{R}$ denotes artery landmark position along the image $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$0-dimension. This is a joint classification-and-regression problem: if the carotid artery is absent, the classification head outputs $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$1 and the regression output is invalid; if the artery is present, the classification head outputs $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$2 and the regression head predicts the normalized position

$\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$3

After pretraining, the fully connected layers are removed and the backbone initializes the ultrasound encoder inside the diffusion-policy system. The same pretrained ResNet is also reused during navigation to output the carotid $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$4-coordinate in pixels, which is central to both the guidance mechanism and the evaluation metrics.

3. Multisensory policy architecture

UltraDP is organized into four modules: data collection, pretraining of an ultrasound encoder, a navigation system based on diffusion policy, and a hybrid force-impedance control system. The learned navigation component consumes four modalities: ultrasound images, wrist camera RGB-D images, contact wrench, and probe pose. The ultrasound stream is encoded by the pretrained and then fine-tuned ResNet features; the wrist RGB-D stream is encoded by a visual backbone; and the pose and wrench are concatenated as low-dimensional state variables into the observation representation $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$5 (Chen et al., 19 Nov 2025).

The core policy is a Denoising Diffusion Probabilistic Model used as a visuomotor controller. Rather than regressing a single deterministic action, it generates actions through iterative denoising conditioned on observation: $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$6 where $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$7 is the denoising step, $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$8 is the current noisy action, $\bm O_t = (\bm U_t, \bm I_t,\bm x_t,\bm w_t),$9 is the learned noise-prediction network, and $\bm U \in \mathbb{R}^{H_1 \times W_1},$0 is Gaussian noise. The action corresponds to the desired next-step pose change plus wrench target, consistent with the demonstration action formulation. The model is trained on observation-action pair sequences with a receding horizon, but the paper explicitly provides the denoising equations rather than a formal training-loss expression.

The stated rationale for diffusion policy is that ultrasound scanning exhibits multi-modal action distributions. When the image is poor, several actions may all be valid: translating the probe, tilting or rotating it, increasing or decreasing pressure, or adjusting torque. A deterministic regressor such as standard behavior cloning tends to average these modes and may produce actions that are not useful. Diffusion policy instead models a conditional action distribution and can generate coherent trajectories that preserve a single mode over time. In the paper’s interpretation, this is especially relevant because the task is organized around image quality and anatomy positioning rather than a fixed Cartesian path.

4. Guidance-augmented denoising and artery centering

A distinctive feature of UltraDP is the specialized guidance module added during inference to bias the policy toward artery-centering actions. The motivation is that expert demonstrations almost always keep the artery centered, but that behavioral regularity is implicit rather than explicitly labeled in imitation data. The guidance mechanism therefore injects a task-specific directional signal into the denoising process (Chen et al., 19 Nov 2025).

During inference, the denoising step is modified to

$\bm U \in \mathbb{R}^{H_1 \times W_1},$1

where $\bm U \in \mathbb{R}^{H_1 \times W_1},$2 is a scaling factor and

$\bm U \in \mathbb{R}^{H_1 \times W_1},$3

is the guidance term. The imaging principle used to construct this guidance is

$\bm U \in \mathbb{R}^{H_1 \times W_1},$4

with $\bm U \in \mathbb{R}^{H_1 \times W_1},$5 a known probe parameter, $\bm U \in \mathbb{R}^{H_1 \times W_1},$6 the probe displacement along the Cartesian $\bm U \in \mathbb{R}^{H_1 \times W_1},$7-axis, and $\bm U \in \mathbb{R}^{H_1 \times W_1},$8 the artery-pixel displacement in the ultrasound image. Because this relation is linear, the gradient with respect to action is constant. The artery location is estimated by the pretrained ultrasound landmark regressor, and the image-space offset from the horizontal center is converted into an action-space bias.

The guidance is therefore a soft directional constraint rather than a hard projection. It does not replace the learned diffusion prior; instead, it nudges denoising toward actions that reduce artery offset while preserving the multimodal structure learned from demonstrations. The paper is explicit that this mechanism is applied during inference rather than as a training-time regularizer.

5. Hybrid force-impedance execution and interaction safety

The navigation policy produces desired pose and wrench targets at about 10 Hz, but physical execution is delegated to a hybrid force-impedance controller running at 1 kHz. The manipulator dynamics are written as

$\bm U \in \mathbb{R}^{H_1 \times W_1},$9

and the control torque is

$\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$0

with main-task torque

$\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$1

and null-space torque

$\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$2

Here,

$\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$3

For the Franka robot used in the system, $\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$4 (Chen et al., 19 Nov 2025).

The controller decomposes pose and force objectives by selection matrices. Force control is assigned along the probe $\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$5-axis and impedance control is assigned in the remaining directions. In the end-effector frame,

$\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$6

$\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$7

and these are transformed to the base frame to obtain $\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$8 and $\bm I \in \mathbb{R}^{H_1 \times W_1 \times 4},$9. This means that normal contact force is actively regulated while lateral position and orientation remain compliant. The null-space term controls redundant degrees of freedom without interfering with the main task.

The paper presents this architecture as the basis of safe interaction. Its concrete safety-related mechanisms are explicit force regulation along the contact-normal direction, impedance behavior in the remaining dimensions, compliant null-space behavior, high-rate execution at 1 kHz, and low-pass filtering of upper-level outputs because the navigation policy runs only at about 10 Hz. The same controller is also used during demonstration collection, with all stiffness and damping parameters set to zero except force control, making the system fully compliant to the sonographer’s movements while assisting contact establishment. Under this assistance mode, expert scans were completed in about 30 seconds, less than half the time needed without assistance.

6. Empirical performance, generalization, and ablations

The hardware platform consists of a Franka manipulator, an ultrasound machine with probe, an ATI Mini40 force/torque sensor between the flange and probe, and an Intel RealSense D405 wrist-mounted camera, with a control computer using an AMD Ryzen 5 5600G CPU, an Nvidia RTX 3060 GPU, and Ubuntu 20.04. The paper evaluates UltraDP first in a simulation-style trajectory setting and then in real-world autonomous scanning on unseen subjects (Chen et al., 19 Nov 2025).

In the trajectory evaluation, observations from demonstrations are fed to the policy and predicted next-step actions are compared to real trajectories. For unknown volunteers, the mean transverse tracking error is $\bm x \in SE(3),$0, mean force is $\bm x \in SE(3),$1, mean torque is $\bm x \in SE(3),$2, and $\bm x \in SE(3),$3. For known volunteers, the corresponding values are $\bm x \in SE(3),$4, $\bm x \in SE(3),$5, $\bm x \in SE(3),$6, and $\bm x \in SE(3),$7. Maximum forces remain below about $\bm x \in SE(3),$8, which the authors interpret as stable and comfortable.

The real-world autonomous evaluation compares UltraDP with Baseline VS, a rule-based visual servoing method modified from Yan et al. with an added image-space servoing term using the pretrained artery regressor, and Baseline BC, a classic behavior-cloning model trained on the same data and using the same low-level controller. Testing is conducted on four unseen volunteers—two female and two male—with five trials per participant, for 20 real-world trials in total. A trial is successful if the ultrasound image shows a clear centered transverse view for more than 60% of the scanning period and the probe stops upon detecting a bifurcation. UltraDP achieves 19/20 successful scans, a mean landmark distance to center of 5.71 px, mean SSIM with human expert videos of 76.83%, a mean expert score of 7.11, and mean $\bm x \in SE(3),$9 of 0.187 N/s. Baseline VS records 17/20, 6.52 px, 71.24%, 6.76, and 0.329 N/s, respectively. Baseline BC records 6/20, 25.44 px, 62.19%, 6.14, and 0.189 N/s.

These results support several narrower conclusions stated in the paper. UltraDP attains the best overall success and artery-centeredness; it exceeds the rule-based visual-servoing baseline despite that baseline having explicit servoing; it substantially outperforms standard behavior cloning in landmark centering and task success; and it maintains smooth force changes, much better than VS and comparable to BC. A modality ablation further shows that both pose and wrench are critical: full UltraDP achieves 19/20, UltraDP without wrench achieves 0/20, UltraDP without pose achieves 8/20, and UltraDP without $\bm w \in \mathbb{R}^6.$0pose, wrench$\bm w \in \mathbb{R}^6.$1 does not converge during training. The qualitative failure modes align with these results: without wrench the model often fails to maintain neck contact, whereas without pose it may maintain contact but lose artery centering and drift out of view. A questionnaire with 24 valid responses indicates that UltraDP is perceived as superior in comfort and efficiency.

7. Limitations, future work, and nomenclature

The explicit future-work statement is brief: “The scaling-up validation will be our future work.” Within the reported scope, the most evident limitation is evaluation scale: real-world autonomous testing is performed on four unseen volunteers over 20 trials. Additional implicit limitations stated in the discussion include the focus on transverse carotid scanning rather than the entire diagnostic workflow, the absence of a formal proof of controller safety or of diffusion-guidance stability, the reliance of centering guidance on the simplified linear imaging relation $\bm w \in \mathbb{R}^6.$2, and the dependence on a pretrained artery detector-regressor (Chen et al., 19 Nov 2025).

A separate nomenclature issue is that UltraDP should not be conflated with UDPNet, whose acronym expands to “Unleashing Depth-based Priors” and denotes a framework for depth-prior-guided image dehazing rather than robotic ultrasound scanning. UDPNet is methodologically centered on DepthAnything V2, the Depth-Guided Attention Module, and the Depth Prior Fusion Module for image restoration, and the paper explicitly notes that “UDP” stands for “Unleashing Depth-based Priors,” not UltraDP (Zuo et al., 11 Jan 2026).

Within its own domain, UltraDP is best understood as a hierarchical, multimodal, force-aware generative control system for robotic ultrasound. Its top level uses a diffusion policy to model ambiguous mappings from image-contact state to action; its inference stage adds task guidance to preserve artery centering; and its execution layer uses hybrid force-impedance control to maintain safe human-robot interaction. The combination is intended to address the specific failure modes of calibration-heavy rule systems and deterministic imitation policies in autonomous carotid scanning.