Implant-Anchored Actions for Robotic Peri-Implant Debridement That Transfers Across Clinics

Peri-implantitis is a bacterial infection around a dental implant that slowly dissolves the bone holding it. It affects over a quarter of implants. Treatment depends on scraping a sticky bacterial film (biofilm) off a screw-threaded titanium surface that sits at the bottom of a narrow, bleeding bone defect, where the clinician often cannot see the thread recesses. Robots that learn from demonstrations are now good at this kind of contact-heavy work in the lab. The most capable of them are diffusion policies, generators that produce robot motions by gradually denoising random noise and can therefore commit to one of several valid strategies instead of averaging them. They usually run inside a two-speed control stack: a slow learned planner and a fast force controller. But medical robot policies are almost always tested on the same robot, cameras and phantoms they were trained on. Clinical prediction models teach a warning here: such models look excellent internally and then degrade at a new hospital, because they learned local shortcuts.

This note proposes that the fix for robots mirrors the fix found for radiograph classifiers. There, forcing the model to attend to the lesion itself, rather than to the whole image, removed hospital shortcuts. For debridement, we propose making the slow diffusion planner speak only in coordinates attached to the implant surface: which thread turn and flank to clean, at what tool angle, in which sweep direction, and with how much pressure. A fast force loop then turns those surface-level intentions into motion on whatever robot is present. The claim is that this change of action space, and not more data or a bigger model, is what lets a policy trained at one site keep working at a different site with a different robot, camera and implant brand. We test that claim with a pre-registered external validation protocol, which to our knowledge no surgical robot policy has been subjected to.

Core Claim

A slow-fast debridement policy whose slow layer outputs implant-anchored contact intentions will show a much smaller drop in performance from internal to external sites than the same architecture outputting Cartesian end-effector poses. This holds when both are trained on identical demonstrations from a single site. We define the transfer gap as $\Delta = S_{\text{int}} - S_{\text{ext}}$, where $S_{\text{int}}$ is the debridement success rate on held-out trials at the training site and $S_{\text{ext}}$ is the success rate at sites whose hardware and phantoms were never seen in training. The prediction is that implant-anchored actions roughly halve $\Delta$ relative to Cartesian actions while matching them internally.

The reasoning is as follows. A Cartesian pose target entangles three things: the task, which is where on the thread to press; the robot, through its kinematics and calibration; and the camera-to-robot registration. Only the first transfers. In implant coordinates the task becomes a small, reusable vocabulary. That vocabulary is still genuinely multi-modal: a clinician may sweep clockwise or counterclockwise along the thread helix, attack the upper or lower flank, or approach around the abutment from the buccal or lingual side. Averaging these choices yields a tool pressed into the thread crest, which is exactly the failure a diffusion policy avoids.

What Prior Work Supplies and Leaves Open

The method is assembled from existing parts. The novelty is where the interface between the two control rates sits and how the system is evaluated.

The two-speed diffusion design comes from three papers, and all three emit actions in robot or world Cartesian space. UltraDP runs a 10 Hz multi-sensor diffusion policy over a 1 kHz hybrid force-impedance controller for carotid ultrasound. Reactive Diffusion Policy pairs a 1–2 Hz latent diffusion planner with a sub-millisecond tactile decoder running above 20 Hz. Compliant Residual DAgger adds a 50 Hz residual policy on top of a 1 Hz base policy. All three work on compliant soft tissue or household objects. None is tested on rigid metal in a confined cavity, and none on hardware other than its own. DIPCOM shows that a diffusion policy can also output stiffness for an impedance controller, and we reuse this idea for pressure intent.

Demystifying Action Space Design motivates our central bet: it shows that the choice of action space changes what diffusion policies can represent. We extend that question from within-setup performance to cross-setup transfer.

The surgical debridement robots are closest in task but use simple controllers. Seita et al. use calibrated positioning with a neural network plus random forest correction on the da Vinci Research Kit. Patel et al. use intermittent synchronization to rhythmic motion. Both are vision-only, lack force feedback, and use pumpkin seeds and raisins as targets. They show that debridement can be automated but say nothing about threads or transfer.

The external-validation literature is what makes this a hypothesis rather than an engineering exercise. Luo et al. found that lesion-level supervision survived external cohorts where image-level supervision did not. The contrast-enhanced mammography study in Radiomic Analysis of CEM and the review in PET/MRI Radiomics show how pooled or single-center evaluation inflates results. Within peri-implantitis itself, the immune-profiling model of Machine learning-assisted immune profiling rests on 24 patients with no external cohort, and the peri-implantitis risk model protocol explicitly calls for multi-center validation.

The same immune-profiling paper also constrains our claims. It finds that the host immune state predicts healing after surgery better than plaque indices do. Better mechanical cleaning is therefore necessary but not sufficient for clinical success, so our endpoints stop at surface decontamination and damage, not healing.

Research Program

Slow layer. The slow layer is a diffusion transformer running at 5–10 Hz. It sees an endoscopic RGB-D image, the implant pose estimated by registering a known implant model (library CAD, or planning CBCT where available) to the image, and a short force history. It outputs a chunk of implant-frame intentions. Each intention is a point on the unrolled thread surface given by helix angle and axial position, a flank choice, a tool tilt relative to the local surface normal, a sweep direction and speed, and a normal-force setpoint.

Fast layer. A 1 kHz admittance controller maps each intention into the robot's joint space using the current implant registration. It regulates normal force against the titanium and caps tangential force and tool jerk. It also absorbs small patient-like motion, simulated by mounting the jaw on a moving platform, through continuous force regulation rather than by waiting for motion extrema. A lightweight residual in the style of CR-DAgger can correct the tangential path from force feedback. This layer holds everything site-specific: robot kinematics, force-sensor bias, and tool compliance.

Demonstrations. Periodontists will teleoperate with a haptic interface at a single site, the training site. They will use a titanium brush and a curette tip on implants in printed jaw phantoms. Defect shapes include circumferential, dehiscence and deep narrow types, with implants coated in a stained biofilm simulant.

Test ladder. The ladder runs from stained-simulant phantoms, to ex vivo porcine jaws with implants placed into surgically created defects, to implants carrying in vitro grown multispecies biofilm. The last step lets us report viable-bacteria counts as well as stain coverage.

Protocol. The external-validation protocol is pre-registered. Two external sites will run frozen checkpoints. Each uses a different arm, force sensor and camera, a different implant brand and thread pitch, and phantoms built independently from a shared written specification. Site staff who did not train the policies will score the outcomes. No tuning is allowed at external sites; only the one-time implant-registration step runs there.

Baselines

Naive reference points. The do-nothing reference is the contamination left after the phantom's standard irrigation step. The strong naive reference is a scripted exhaustive helical sweep. It plans a fixed toolpath along every thread turn from the registered CAD model and executes it through our same 1 kHz force loop. This baseline tests whether learning adds anything once geometry is known. We also report a clinician using hand instruments on identical phantoms as the human reference.

Strongest existing method. The strongest competitor is an UltraDP-style system: a multi-sensor diffusion policy emitting Cartesian waypoints over a hybrid force-impedance controller, re-implemented from UltraDP. It is the published system closest to ours in both architecture and medical contact. On its own unseen-subject test it reports 95% success, against 85% for rule-based visual servoing and 30% for behavior cloning. No reported result was found on Lacuna for any method on peri-implant or dental implant debridement, so all numbers on our task will be new. We also run a Reactive Diffusion Policy variant, because its fast tactile correction is the most reactive design in the file. It reports 80% success under perturbation versus 15% for open-loop chunks.

Unimodal baseline. A Gaussian behavior-cloning policy and an action-chunking transformer, both using our implant-frame action space, test whether multi-modality matters here.

Fairness rules. All learned methods get the same demonstrations, the same observations, the same fast controller wherever the method allows it, the same GPU-hour budget, and the same 20-configuration tuning budget. Tuning uses only training-site validation trials.

Ablations. Two ablations isolate the key ingredient. First, our system with Cartesian instead of implant-frame outputs, holding everything else fixed. Second, our system with the fast force loop replaced by stiff position tracking of open-loop chunks.

What counts as winning. Our system beats the UltraDP-style baseline if two conditions hold. Its external success must be significantly higher, by a pre-registered mixed-effects test with site and phantom as random effects. Its transfer gap $\Delta$ must be at most half of the baseline's, with no increase in damage events. Damage is defined as peak normal force above a pre-set threshold, measurable thread scoring under profilometry, or crest fracture in the bone-mimic layer. It beats the scripted sweep if it achieves equal or lower residual contamination in occluded recesses in less time, again with no more damage.

Why not just…?

Why not just plan a helical toolpath from the implant model and skip learning? This is the most serious threat, and it is run as a first-class baseline for exactly that reason. A known thread geometry plus force control may clean the exposed, visible surface as well as any learned policy. The case against it is specific to peri-implant defects. The reachable tool angle changes with the shape of the bone wall. Some thread segments are occluded and need re-orientation or a switch in approach side. Contamination is patchy, so exhaustive sweeping wastes time and adds contact exposure. The only evidence in the file is indirect: rule-based visual servoing trailed UltraDP by ten points with jerkier forces, and continuous tracking reached only 62% in Patel et al.'s debridement task. If the scripted sweep matches us on deep, narrow defects, we will report that learning is unnecessary for this task and that the useful contribution is the force loop and the validation protocol.

Why not just use an UltraDP-style Cartesian diffusion policy, since it already generalizes to unseen subjects? Its generalization was measured on new people scanned with the same robot, probe and room. Our Cartesian-output ablation tests whether that survives a hardware change. If Cartesian and implant-frame policies show similar transfer gaps, the central hypothesis is false.

Why not just collect demonstrations at all three sites and pool them? This is a real alternative, and pooling may shrink the gap. However, Luo et al. found that adding external data without better supervision did not remove shortcuts, and pooled splits hide the problem entirely. We therefore run a leave-one-site-out pooled Cartesian policy and compare it with our single-site implant-frame policy on the held-out site. If pooling wins, the practical recommendation becomes multi-site data collection, and we will say so.

Why not just use Gaussian behavior cloning or an action-chunking transformer, which are cheaper? In the file's comparisons, unimodal or VAE-based policies lose badly in contact tasks with competing strategies: behavior cloning reached 30% in UltraDP, and Comp-ACT produced 9.96% fine powder versus 55.88% for DIPCOM in powder grinding. Thread cleaning has discrete sweep-direction and flank choices. We will measure directly how often unimodal policies drive the tool into the thread crest, the signature of averaging two strategies.

Decisive Checks

The first check is the transfer gap. Across the two external sites, implant-anchored actions should cut $\Delta$ by at least half relative to Cartesian actions. The absolute gap for Cartesian actions tells us whether a real problem existed to solve at all.

The second is the force-loop ablation. Removing the fast loop should raise damage events on rigid titanium markedly. If it does not, the two-speed design is unnecessary at this scale.

The third check is the scripted-sweep comparison on occluded recesses and treatment time. It decides whether a learned slow layer is justified.

The fourth is the ladder from phantom to ex vivo to live biofilm. The ranking of methods on stain coverage should agree with the ranking on viable-bacteria counts. Otherwise stain is a misleading surrogate, and future work should not report it alone.

A negative result on any of these is informative and will be published with the pre-registered protocol and all site data.