An RL-fine-tuned small model lets a robot follow ultrasound guidelines to scan gallbladder, spine, and kidney autonomously

It introduces a unified, guideline-driven agentic robotic ultrasound framework in which an LLM acts as a high-level planner, retrieving guideline steps from scanning handbooks and dynamically invoking perception and control tools, thereby supporting variable, decision-dependent workflows such as repeating steps, adapting scanning strategies, and adjusting trajectory or force based on image quality. Unlike prior rule-based systems that explicitly encode expertise as fixed procedures and task-specific models, this framework does not hard-code scanning steps; the agent re-reasons at each step from observations and the current state to choose the next tool call. The paper details three components (robotic/perception tools, a guideline database, and an LLM as high-level planner) and five tool types: trajectory planning, robot execution, contact adjustment, voice guidance, and trajectory refinement; the evidence comes from method description plus subsequent experiments, not conceptual argument alone.

It uses reinforcement-learning-based fine-tuning (SFT followed by PPO, with LoRA ranks of 16 and 8) to improve a small model's planning-oriented reasoning and tool-calling behavior, with a dense reward combining tool-name matching, argument-key presence, argument-value correctness, and a penalty for spurious extra arguments. Prior LLM-for-robotic-ultrasound work typically emphasizes guideline execution and tool use, with the agent's intermediate reasoning less explicitly modeled or evaluated; this work optimizes reasoning together with the correctness of tool selection and parameterization. The reward is given as equations: an incorrect tool name yields −0.5, otherwise clip[0,1](0.1 + 0.9·s_args), where s_args mixes key presence, value correctness (exact matching for discrete fields, tolerance-based matching for selected numeric fields), and an extra-key penalty.

Fine-tuning Qwen3-4B-Instruct-2507 on 10 training guidelines and evaluating on 10 held-out guidelines with no content overlap raised step-wise accuracy from 0.6512 to 0.8973 and overall success rate from 0.5384 to 0.9230. The evaluation targets unseen guideline content, testing generalization to unseen guideline instructions rather than reproduction of training guidelines. Reference outputs were generated by Qwen3-30B-A3B-Instruct-2507, and the tool set stayed identical across training and testing; results are reported in a table and constitute verbal-level execution validation.

Real-world robotic scanning was performed on three human volunteers for gallbladder, spine, and kidney; all spine and kidney experiments were successful, and one gallbladder case failed because of higher procedural complexity requiring breath-hold instructions and close patient-robot coordination. The work moves the same unified pipeline from verbal validation to a real-world feasibility demonstration across multiple anatomical targets, rather than a dedicated system for a single anatomy. Experiments used a KUKA LBR iiwa robot with a Siemens ACUSON Juniper ultrasound system and a 5C1 probe in a 3D-printed holder, an Epiphan frame grabber, and a Microsoft Azure Kinect RGBD camera on the end effector; task success was judged by whether the target organ was correctly visualized in the ultrasound image, with IRB approval and written informed consent from all participants.

The framework targets settings where autonomous ultrasound should follow clinical scanning guidelines, on robotic platforms with a manipulator, an ultrasound system, an RGBD camera, and accessible GPU hardware; it is designed so one system follows different guidelines and covers multiple anatomical targets within a unified pipeline, with verbal validation on 10 unseen guidelines and real-world validation on gallbladder, spine, and kidney. For a reader, this means the work offers a reusable guideline-to-action planning and tool-calling paradigm plus an interface for inspecting reasoning traces, not a clinically deployable product.

Real-world evidence comes from three volunteers across three anatomical targets, and one gallbladder case failed, so stability across more anatomies and subjects remains an open question; verbal evaluation uses 10 training and 10 held-out guidelines, a limited number and coverage, so behavior when extending to more guidelines and more complex interactive workflows is still to be observed. In addition, results are reported in tables and figures, so readers of the text alone may need to consult the original figures for details such as specific tool parameter values and the full course of the failed case.