TL;DR
RobOralScan is the first reinforcement learning (RL) framework designed for automated robotic intraoral dental scanning. By combining a tri-state geometric memory with tooth-wise coverage rewards, it achieves high-quality, uniform 3D reconstructions of teeth and transfers zero-shot to physical robot arms.
What Problem Does This Paper Address?
Getting a digital 3D model of a patient's teeth (an intraoral scan) is a foundational step in modern dentistry for crowns, implants, and braces. However, doing a full-arch scan manually is a high-burden, labor-intensive task. The human mouth is a highly confined space, and intraoral scanners have an incredibly narrow field of view (FOV). Operators must continuously adjust the scanner's 6-DoF pose, distance, and speed to avoid missing surfaces. Consequently, scan quality varies wildly based on operator experience and workload.
While robotic automation has been explored in medical imaging fields like ultrasound, automating dental scanning is uniquely challenging. A robot must dynamically plan a path that ensures complete, even coverage across every individual tooth while navigating tight spatial constraints.
Methods
RobOralScan formulates intraoral scanning as a feedback-driven, closed-loop reinforcement learning problem optimized via Proximal Policy Optimization (PPO). Instead of executing a rigid, pre-planned path, the policy determines incremental, relative 6-DoF scanner motions based on the current state of the reconstruction.
The architecture relies on two core innovations:
- Geometric Memory-Based Observation Space: Rather than looking only at the current video frame, the policy observes a volumetric grid that accumulates past scan data. Using 3D ray-casting, voxels are classified into a tri-state representation: occupied (scanned surface), free (empty space traversed by the scanner's light rays), or unknown (unexplored space). This allows the policy to explicitly reason about unexplored frontiers.
- Tooth-Wise Coverage Learning: Standard RL reward functions typically maximize total surface coverage. However, this causes agents to over-scan easy areas (like the flat tops of molars) while neglecting tricky side surfaces. RobOralScan introduces an equity reward that prioritizes the least-scanned teeth (focusing on the lower-tail mean of the tooth coverage distribution) and penalizes the robot if it leaves a tooth prematurely. This is paired with a progressive curriculum training scheme that gradually increases coverage requirements over time.
Datasets
The authors built a synthetic scanning environment using the Teeth3DS dataset, which provides 3D intraoral scan models complete with per-tooth segmentation labels.
- Training: 80 jaw models.
- Evaluation: 10 held-out models from unseen patients.
Note: Ground-truth tooth annotations were exclusively used to calculate rewards during training and metrics during evaluation; they were never provided as direct inputs to the policy network.
Key Insights and Ideas
- The Global Coverage Trap: Optimizing purely for global coverage is insufficient for dental applications. A baseline model trained only on global coverage achieved a decent 71.89% average coverage but left individual teeth completely missed (scoring a dismal 2.61% in lower-tail per-tooth coverage). Tracking per-tooth equity is mandatory.
- The Power of "Unknown" Space: Explicitly separating "free space" from "unknown space" provides critical guidance. When the authors replaced their tri-state memory with a standard binary occupancy grid (occupied vs. unoccupied), the agent's success rate plummeted from 8/10 to 5/10 because it could no longer identify genuinely unexplored frontiers.
- Sim-to-Real Viability: Despite being trained entirely in a simulated environment, the policy learned robust geometric principles that allowed it to transfer zero-shot to physical hardware without real-world fine-tuning.
Key Results
RobOralScan significantly outperformed classical Next-Best-View (NBV) algorithms and scripted anatomy-guided baselines:
- High Coverage: Achieved an average final coverage of 92.58% and a lower-tail per-tooth coverage of 88.45%.
- High Accuracy: Maintained a remarkably low Chamfer Distance of 0.00838.
- Reliability: Successfully met the strict clinical scan completion criteria in 8 out of 10 evaluation episodes, whereas all baseline methods scored 0 out of 10.
- Physical Deployment: The system was validated on a real-world hardware setup consisting of a Franka Research 3 robot arm and a Huvitz Lilivis SCAN intraoral scanner, demonstrating stable trajectories and progressive reconstruction on a dental phantom.
Limitations and Future Work
The primary limitation is that the physical hardware validation was conducted on a rigid plastic dental phantom rather than inside an actual human oral cavity. Real-world clinical deployment introduces complex variables such as patient movement, soft-tissue interference (tongue and cheeks), saliva moisture, and stricter spatial access limits.
Additionally, discrepancies between the robot's commanded actions and its actual executed movements can introduce pose uncertainty, causing errors to propagate into the geometric memory. The authors note that future work will focus on integrating online registration and closed-loop pose correction using real-time scanner feedback to make the system resilient enough for live patient trials.
Summary updated