IRIS: Intelligent Robotic Imaging System

• IRIS is a versatile robotic imaging system that co-develops mechanical design, sensing, control, and learning for both ultrasound and cinematic applications.
• It features a 6-DOF architecture capable of precise probe positioning, compliant force control, and goal-conditioned visuomotor execution in a low-cost setup.
• The platform demonstrates dual usage by combining a blueprint for robotic sonography with a dedicated cinematic robot arm for autonomous and repeatable motion.

Searching arXiv for the specified IRIS-related papers and surrounding context. Intelligent Robotic Imaging System (IRIS) is an acronym used in recent arXiv literature to denote two closely related but distinct classes of robotic imaging systems: a blueprint for intelligent robotic ultrasound imaging systems reviewed by Bi et al., and a task-specific 6-DOF robotic camera arm for autonomous cinematic motion control. In both usages, IRIS denotes a tightly integrated stack in which mechanism design, sensing, control, perception, and learning are co-developed rather than treated as separable subsystems. The ultrasound formulation emphasizes robotic sonography, compliant contact, signal-to-image processing, and machine learning–driven autonomy; the cinematic formulation emphasizes low-cost mechatronics, goal-conditioned visuomotor imitation learning, and autonomous execution of camera trajectories (Bi et al., 2024, Cheng et al., 19 Feb 2026).

1. Scope, nomenclature, and domain usage

In the available literature, IRIS is not restricted to a single device class. One usage describes a generalized architecture for robotic ultrasound, while another names a specific manipulator for cinema robotics.

IRIS usage	Domain	Core characteristics
Intelligent robotic imaging system blueprint	Robotic ultrasound	6 DOF probe positioning plus at least 1 DOF force; compliant control; ML-driven autonomy
IRIS platform	Cinematic robot arm	Purpose-built 6-DoF manipulator; ACT-based goal-conditioned imitation learning; cost ≈\$950; payload 1.5 kg

This dual usage matters because the common acronym can obscure a substantive difference in epistemic status. In the ultrasound setting, IRIS denotes a system-level blueprint assembled from commonly employed robotic mechanisms, control techniques, image-processing stages, and machine learning modules. In the cinematic setting, IRIS denotes a concrete hardware-software platform with specified kinematics, bill of materials, training protocol, and experimental evaluation. A common misconception is therefore to treat IRIS as a single standardized platform; the literature instead uses the term for both a conceptual architecture and a named embodiment (Bi et al., 2024, Cheng et al., 19 Feb 2026).

2. Robotic ultrasound IRIS: mechanical architecture and sensing

For robotic ultrasound, serial manipulators such as 6- or 7-DOF industrial arms are described as the most common mechanism because of their large workspace and ease of programming, whereas parallel robots in Stewart-platform style offer higher stiffness and more compact footprints but reduced reach. Some groups add passive compliance elements, including spring-loaded clutches, soft polymer mounts, or flexible rails, to limit excessive contact force and improve patient safety (Bi et al., 2024).

A fully actuated IRIS in this setting usually provides 6 DOF for probe positioning, comprising 3 translations and 3 rotations, plus at least 1 DOF for controlled contact force along the beam axis. Redundant DOFs, specifically 7 or 8 joints, facilitate impedance control in the null space of the primary task, improving compliance without sacrificing trajectory tracking. This mechanical specification reflects the dual requirement of anatomical reach and stable acoustic coupling, which distinguishes robotic ultrasound from free-space manipulation.

The end-effector is not merely a probe mount but an instrumented interface. Motorized probe clamps with integrated 6-axis force/torque sensors at the wrist enable closed-loop force control, while soft conformal holders or pneumatically attachable rails passively conform to anatomy. Auxiliary sensing may include IMUs for probe orientation, external RGB-D or stereo cameras for surface registration and motion tracking, and optical markers for patient registration. The ultrasound machine itself provides raw RF streams, B-mode streams, Doppler data, and internal confidence maps. Taken together, these components define a multimodal sensing substrate in which contact mechanics, probe pose, anatomy, and acoustic signal quality are jointly observable (Bi et al., 2024).

3. Control and signal-to-image pipeline in ultrasound IRIS

The control layer in robotic ultrasound spans classical model-based formulations and more recent adaptive or learning-assisted approaches. Hybrid position/force control is described as a Cartesian-space velocity law in the free directions with force feedback along the imaging axis. Impedance or admittance control prescribes a desired dynamic relationship,

$$M_d \left( \ddot{x} - \ddot{x}_{ref} \right) + B_d \left( \dot{x} - \dot{x}_{ref} \right) + K_d \left( x - x_{ref} \right) = F_{ext}.$$

Computed-torque control is given by

$$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$

where $q$ denotes joint angles, $M$, $C$, and $G$ denote the inertia matrix, Coriolis/centrifugal and gravity terms, $\tau$ denotes actuator torques, and $J^T F_{ext}$ is the external force transformed to joint space. Modern variants include variable impedance via online stiffness estimation, for example with the Hunt-Crossley contact model, hierarchical QP control that prioritizes force objectives and resolves orientation in the null space, event-triggered adaptive backstepping for force/position tradeoff, and learning-based controllers that tune gains via Bayesian optimization or deep kernel regression (Bi et al., 2024).

The imaging pipeline begins with raw data capture: the ultrasound system digitizes RF echo signals from each transducer element. Beamforming is represented by delay-and-sum reconstruction,

$$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$

where $r_i$ is the received signal at element $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$0, $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$1 is the round-trip delay to pixel $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$2, and $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$3 are apodization weights. Demodulation and envelope detection then proceed through Hilbert transform, envelope extraction, log compression, and scan conversion to B-mode. For volumetric imaging, 2D-to-3D reconstruction is performed either by free-hand acquisition, in which probe pose is tracked via encoders or camera and successive B-scans are re-sampled into a volumetric grid, or by mechanical sweep, in which an actuated linear stage moves the probe along a fixed path while recording frames.

Post-processing includes speckle and noise reduction via anisotropic diffusion or non-local means, confidence map estimation with a random-walk model to quantify acoustic shadowing and signal reliability, and segmentation and registration modules that feed back into control for task-driven probe repositioning. This arrangement makes the signal-processing pipeline operationally inseparable from the controller: image quality estimation and anatomical localization directly influence the next probe action (Bi et al., 2024).

4. Machine learning autonomy, data representation, and open problems in ultrasound IRIS

Bi et al. organize machine learning–driven autonomy into implicit interpretation and direct reasoning. Under implicit interpretation, modular perception and control are coupled through intermediate representations. CNN-based segmentation with U-Net, Mask R-CNN, or W-Net is used to locate vessels, organs, or lesions, with typical losses

$$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$4

and

$$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$5

Registration networks or classical ICP on point clouds align patient anatomy with pre-operative CT or MR templates for trajectory planning, and segmentation outputs can reinforce control laws, for example through centerline extraction followed by lateral guidance. Under direct reasoning, reinforcement learning in a POMDP formulation uses state equal to the ultrasound image with or without a depth camera, action equal to $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$6, and a reward $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$7 that combines image confidence $$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$8, proximity to a standard plane, and force safety, with objective

$$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + J^T(q)F_{ext},$$9

Learning from demonstrations is described through behavior cloning,

$q$0

and max-entropy inverse reinforcement learning,

$q$1

Temporal or spatial ranking losses are used to cope with suboptimal expert traces. Hybrid architectures integrate learned perception modules into a rule-based planner for predictable behavior (Bi et al., 2024).

Data scarcity is treated as a central limiting factor. Physics-based simulators such as k-Wave and convolution+ray-tracing generate labeled 2D or 3D ultrasound from CT, MR, or scatterer maps, while GAN and CycleGAN models create synthetic volumes and align appearance with real ultrasound domains. Data augmentation includes standard geometric transforms and intensity shifts, as well as physics-inspired perturbations that simulate deformation, reverberation, and variable SNR to respect ultrasound physics. Representation learning methods include domain disentanglement via adversarial loss with a feature extractor $q$2 and discriminator $q$3, mutual information minimization between domain code $q$4 and task code $q$5,

$q$6

and contrastive or clustering losses to align cross-domain features. Multi-modal encoding fuses RF, B-mode, Doppler, and kinematics into a single latent representation using attention or transformer blocks (Bi et al., 2024).

Open problems are identified at the levels of generalization, adaptation, physics integration, and governance. Inter-patient and inter-device variability remain barriers for robust segmentation and control; physics-aware network design such as CACTUSS and LOTUS, together with meta-learning, may help. Real-time adaptation requires patient motion compensation through on-the-fly registration and trajectory replanning, as well as active re-acquisition strategies for Doppler signals or confidence re-estimation. Integration of physics and learning is linked to differentiable simulators and end-to-end training, including Ultra-NeRF, to better capture wave propagation and anatomy. Ethics, safety, and regulation are framed in terms of clear autonomy levels from no autonomy to full autonomy with ISO/IEC guidance through IEC/TR 60601-4-1 and transparent decision logic. A plausible implication is that progress toward full autonomy depends less on isolated perception gains than on calibrated co-design across contact mechanics, acoustic physics, control hierarchy, and regulatory interpretability (Bi et al., 2024).

5. Cinematic IRIS: manipulator design, materials, calibration, and repeatability

The 2026 IRIS platform is a purpose-built 6-DoF robotic arm designed from the ground up for autonomous, learning-driven cinematic camera motion. Rather than adapting heavy industrial manipulators, it co-designs lightweight hardware and visuomotor imitation learning to execute smooth push-in, crane, and dolly-style shots with sub-millimeter repeatability, learn directly from human cinematographers via demonstration, and remain compact and portable for tabletop and small-studio use. Its stated design goals are cost under \$1,000 USD for all mechanical parts, electronics, camera, and compute; payload up to 1.5 kg; total mass approximately 8.5 kg; and sub-millimeter repeatability across repeated trials (Cheng et al., 19 Feb 2026).

The manipulator architecture is a serial 6-DoF arm with decoupled translational joints 1–3 and rotational joints 4–6. Joint 1 provides 360° continuous base yaw; joint 2 provides shoulder pitch of $q$7; joint 3 provides elbow pitch of $q$8 and is belt-driven to relocate the motor proximally; wrist pitch and roll are implemented as a two-motor differential to eliminate motors at the distal tip; and joint 6 provides end-effector yaw of $q$9. Link geometry uses two $M$0 mm carbon-fiber tubes with total reach of approximately 940 mm. The actuators are Unitree GO-M8010-6 brushless DC motors with 6.33:1 planetary reduction, peak torque 23.7 N·m, and top speed 30 rad/s. Twenty-four structural parts, including joint housings, timing-belt pulleys, and mounts, are printed in PLA or PETG, with reported tensile strength of approximately 60 MPa and stiffness sufficient for camera motions up to 3.3 m/s and accelerations to 15 m/s² (Cheng et al., 19 Feb 2026).

The bill of materials includes six Unitree actuators at \$M$150 in filament, belts, pulleys, and fasteners at \$M$2200, Jetson Nano or laptop compute amortized at \$M$3100, for a total cost of approximately \$M$4K = 10$M$5$M$6$

Reported scatter is 0.25 mm at the start waypoint, 0.15 mm at WP2, 0.04 mm at WP3, and 0.20 mm at WP4, with maximum 0.66 mm. This measurement grounds the claim of approximately 1 mm repeatability in a repeated-trajectory setting (Cheng et al., 19 Feb 2026).

The sensing and preprocessing pipeline uses an Intel RealSense D435 RGB camera at 30 Hz and resolution $M$7, with depth available but unused for the imitation-learning policy, and 15-bit absolute joint encoders at 200 Hz via an RS-485 bus. Images are resized and normalized to $M$8 using ImageNet mean and standard deviation, while joint positions and velocities are timestamped and synchronized with images in a sliding-window buffer of length $M$9. Camera-to-robot extrinsics are estimated by a two-step calibration procedure consisting of homing the arm in a vertical zero-torque pose to set encoder origins, followed by hand-eye calibration using a known fiducial grid and standard Tsai-Lenz or PnP routines (Cheng et al., 19 Feb 2026).

6. Goal-conditioned visuomotor learning, autonomous execution, and evaluation in cinematic IRIS

The cinematic IRIS formulates camera motion as a goal-conditioned POMDP. At time $C$0, the observation is

$C$1

where $C$2 is the current RGB frame and $C$3 are joint angles. A goal image $C$4 encodes the desired framing, and the policy predicts an $C$5-step sequence of future joint configurations,

$C$6

with latent style variable $C$7, $C$8, $C$9, and $G$0. The policy architecture combines Action Chunking with Transformers and a CVAE. Its visual encoder is a frozen ResNet-18 backbone followed by spatial softmax and 2D coordinate features, with 11.2 M parameters; the proprioceptive encoder maps 6D joint angles and velocities through a small MLP to a 64D embedding; the CVAE style encoder has 105 k parameters; the transformer encoder and decoder each have 4 layers, model width $G$1, 8 heads, and 7.4 M transformable parameters; and the output head projects decoder outputs to $G$2 joint positions. The standard attention operator is stated as

$G$3

The training objective comprises imitation or reconstruction loss $G$4, latent regularization $G$5, and a smoothness term $G$6 that penalizes jerk in joint space. Their explicit forms are

$G$7

$G$8

and

$G$9

Human experts teleoperate the zero-torque arm to perform push-in shots tracking a coffee cup under unobstructed and obstacle-avoidance conditions. ROS records joint states at 200 Hz and RGB at 30 Hz, and episodes are segmented into sliding windows of length $\tau$0. The dataset totals 132 episodes and 13,954 clips, split 80/10/10 at the episode level, with no additional data augmentation beyond random train/validation splitting. Training uses a single NVIDIA RTX 4090 GPU, batch size 64, AdamW with initial learning rate $\tau$1 and weight decay $\tau$2, for 100 epochs corresponding to approximately 34,500 gradient steps and about 8 hours total, with $\tau$3 and $\tau$4 (Cheng et al., 19 Feb 2026).

At deployment, a ROS node running at 10 Hz collects $\tau$5 and maintains an $\tau$6 buffer. The policy produces $\tau$7 with $\tau$8; the controller then executes a receding-horizon strategy by taking $\tau$9, clamping joint change to at most 0.2 rad, and applying an exponential moving average with $J^T F_{ext}$0 for jitter suppression. Commands are sent to a 200 Hz impedance loop with an $J^T F_{ext}$1-filter of $J^T F_{ext}$2 and velocity limit of 0.6–1.0 rad/s. Inference latency is reported as less than 10 ms on an RTX 3070.

Evaluation includes low-level control without vision and high-level shot tasks with vision and imitation learning. For a circular trajectory over two cycles, low-level control yields tracking RMSE $J^T F_{ext}$3 cm and maximum error 2.10 cm. High-level evaluation covers unobstructed push-in and push-in with a $J^T F_{ext}$4 cm obstacle, with 10 trials per condition from random initial poses. Metrics are visual alignment $J^T F_{ext}$5, defined as cosine similarity between penultimate ResNet features of the last and goal image; success rate, defined as trials finishing without collision and $J^T F_{ext}$6; Cartesian smoothness $J^T F_{ext}$7, defined as the $J^T F_{ext}$8 norm of end-effector jerk; framing error $J^T F_{ext}$9, defined as the pixel distance between image center and object centroid from YOLOv8; subject retention rate; and latency. Mean results over 10 trials are: ACT-CVAE with success 90%, $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$0, jerk $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$1 m/s³, framing error $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$2 px, SRR $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$3%, latency approximately 9.2 ms; Human Replay with success 90%, $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$4, jerk $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$5 m/s³, framing $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$6 px, SRR $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$7%; and RRT* planner with success 10%, $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$8, jerk $$s(\mathbf{x}) = \sum_{i=1}^{N} w_i \cdot r_i(t_i(\mathbf{x})),$$9 m/s³, framing $r_i$0 px, SRR $r_i$1%. The paper further states that on novel unseen initial poses and obstacle geometries, the closed-loop imitation-learning policy consistently re-centers the target and avoids collisions, whereas the open-loop planner frequently collides or misses the framing goal (Cheng et al., 19 Feb 2026).

A comparative analysis situates this platform against commercial cinema robots and 3D-printed arms. Commercial systems such as Bolt, Colossus, and Kira are summarized as costing \$r_i$2\pm 0.05$r_i$31–5 k but having limited repeatability of $r_i$4–5 mm, payload of 0.5–3 kg, and speed below 0.5 m/s. IRIS is reported to achieve reach 940 mm, speed 3.3 m/s, acceleration 15 m/s², repeatability $r_i$5 mm, payload 1.5 kg, and cost approximately \$0.95 k. The stated limitations are payload limited to 1.5 kg, structural flex under high torque, single-camera end effector without multi-view or stereoscopic capability, and an imitation-learning dataset covering only push-in shot style and a single object type. Proposed improvements include stiffer materials or mixed-machining, higher-torque actuators or gearboxes for 3–5 kg payloads, addition of a second camera or wide-angle lens, and dataset expansion to pans, tilts, and multi-goal sequences. This suggests that the cinematic IRIS is best understood not as a general-purpose film robot, but as a tightly scoped demonstration of low-cost, learning-driven visuomotor control under realistic mechatronic constraints (Cheng et al., 19 Feb 2026).