A longer autonomous sequence
Johns Hopkins researchers reported SRT-H, the Surgical Robot Transformer-Hierarchy, in Science Robotics on July 9, 2025. The system performed the clipping and cutting phase of a cholecystectomy on ex vivo pig gallbladders. That phase was divided into 17 tasks, including identifying anatomy, grasping tissue, placing clips and cutting selected structures.
This was a larger challenge than an isolated needle pass or suture. Errors can compound across a long sequence: a weak grasp changes the scene for the next step, and a misplaced tool can make a later action unsafe. SRT-H completed trials on eight previously unseen gallbladders, with the project page reporting success in all eight.
The experiment used realistic biological tissue, but it was not surgery on a living patient. Ex vivo tissue does not reproduce bleeding, breathing, changing physiology or the full range of anatomy and complications in a clinical operating room. The work is a laboratory proof of concept and has not been authorized to operate autonomously on people.
A hierarchy of language and motion
SRT-H uses two learned levels. A high-level policy expresses the current task or a correction in language. A low-level policy converts the selected instruction and camera observations into robot trajectories. The high-level layer also produces a correction flag that switches the controller from the planned task to a recovery instruction when it detects a suboptimal state.
The researchers collected about 18,000 demonstrations across more than 30 pig gallbladders. The dataset included ordinary demonstrations and deliberately created failure states followed by expert recoveries. This taught the policy what to do after a missed grasp, an overshot clip position or another non-catastrophic error, rather than assuming every previous action was perfect.
The system could also accept spoken corrections such as moving one arm in a specified direction. That interface is significant because future autonomy does not have to mean excluding the surgical team. A clinician needs a fast, understandable way to guide or halt a system, and corrections can become additional training examples.
What the tests establish
The project reports that SRT-H completed the clipping and cutting task on all eight test gallbladders without human intervention during execution. It also handled introduced visual changes, including blood-like dye, and shifts in initial tool position. Ablation tests examined the contribution of wrist cameras, corrective instructions, online language tuning and the hierarchical design.
Eight successful specimens provide evidence that the method can generalize beyond its training examples, but the sample remains small. A 100 percent result in eight trials does not imply a 100 percent clinical reliability rate. Rare anatomy, equipment faults and emergencies require far larger and more varied evaluation. The robot also took longer than a human expert.
Clinical autonomy remains a separate threshold
The U.S. Food and Drug Administration describes marketed robotically assisted surgical systems as computer-assisted devices controlled directly by a trained surgeon. SRT-H investigates a different level of autonomy. Before such a system could be used clinically, developers would need controlled studies, robust detection of conditions outside its competence, sterile and validated hardware, cybersecurity controls, clear responsibility for interventions and regulatory authorization.
The study's main contribution is a mechanism for long-horizon recovery. Earlier scripted systems could be precise when the world matched their plan, but biological tissue never presents exactly the same scene twice. A policy that recognizes and repairs small mistakes is a necessary capability for autonomous surgery.
Necessary is not sufficient. SRT-H automated one demanding phase on prepared specimens, not an entire operation on patients. Its results narrow a research gap between short robotic skills and procedure-level behavior while leaving the clinical evidence, supervision design and safety case as future work.