Autonomous Robotic Drilling System for Mice Cranial Window Creation: An Evaluation with an Egg Model

TL;DR

The paper tackles autonomous robotic drilling for mouse cranial window creation by addressing variability in skull geometry with image-based feedback. It introduces a trajectory planner that uses constrained cubic spline interpolation to modulate drilling depth in real time according to a per-point completion map predicted from microscopic images. A DSSD-inspired multi-task network predicts the drilling area and a pixel-wise completion map to enable autonomous adjustment of the drill path, trained on an eggshell surrogate to mimic skull properties. In 20 egg-trial experiments, the system achieved an 80% success rate with an average drilling time of about 16.8 minutes, demonstrating feasibility and indicating a path toward application in actual mouse skull procedures with further refinements.

Abstract

Robotic assistance for experimental manipulation in the life sciences is expected to enable precise manipulation of valuable samples, regardless of the skill of the scientist. Experimental specimens in the life sciences are subject to individual variability and deformation, and therefore require autonomous robotic control. As an example, we are studying the installation of a cranial window in a mouse. This operation requires the removal of the skull, which is approximately 300 um thick, to cut it into a circular shape 8 mm in diameter, but the shape of the mouse skull varies depending on the strain of mouse, sex and week of age. The thickness of the skull is not uniform, with some areas being thin and others thicker. It is also difficult to ensure that the skulls of the mice are kept in the same position for each operation. It is not realistically possible to measure all these features and pre-program a robotic trajectory for individual mice. The paper therefore proposes an autonomous robotic drilling method. The proposed method consists of drilling trajectory planning and image-based task completion level recognition. The trajectory planning adjusts the z-position of the drill according to the task completion level at each discrete point, and forms the 3D drilling path via constrained cubic spline interpolation while avoiding overshoot. The task completion level recognition uses a DSSD-inspired deep learning model to estimate the task completion level of each discrete point. Since an egg has similar characteristics to a mouse skull in terms of shape, thickness and mechanical properties, removing the egg shell without damaging the membrane underneath was chosen as the simulation task. The proposed method was evaluated using a 6-DOF robotic arm holding a drill and achieved a success rate of 80% out of 20 trials.

Figures (7)

• Figure 1: The system setup used in this work, consisting of one of the robotic arms of our AI-robot platform for scientific exploration marques2022design. For this work, we used the arm with a micro drill as an end effector, a 4K camera, and a clamping device for the egg.
• Figure 2: Block diagram of the autonomous robotic drilling system. The robotic system interacts with the egg and a 4K camera provides images to the completion level recognition block. The completion level recognition block outputs the depth of the points for the trajectory planning block. Lastly, the trajectory planning block uses the image information to update the trajectory and sends setpoints to the robotic system, closing the loop.
• Figure 3: Generated curve with $n=30$ points along the circular path. The blue points are the sampling points, the red line is the result of traditional spline interpolation, and the green line is the result of constrained spline interpolation. We can see the red line overshoots while the green does not.
• Figure 4: The overall network architecture for the drilling area detection and completion level prediction. On the left, ResNet-50 is applied as a feature extractor. It is followed by the downscale-stream and the upscale-stream, which consist of a sequence of deconvolution layers interleaved with conv-skip connection. The localization and classification of bounding boxes (top) and pixelwise completion level map (bottom) are performed in a multiscale fashion by single convolutional layers operating on the output of deconvolution layers.
• Figure 5: Output image of the network. (a) Result of bounding box detection; (b) The completion level map; (c) The heatmap of completion level map for visualization; (d) The progress bar for visualization.