DesignAlgorithm_SnakeRaven

Automated end-to-end design optimisation for patient-specific snake-like surgical manipulators, searching over dexterity, reachability and task-space obstacles to generate a bespoke 3D-printable design from a patient scan.

Status: archived. Research code as published in 2022, not actively maintained. Developed against MATLAB R2018b.

Paper

This is the companion code for:

A. Razjigaev, A. K. Pandey, D. Howard, J. Roberts and L. Wu, "End-to-End Design of Bespoke, Dexterous Snake-Like Surgical Robots: A Case Study With the RAVEN II," IEEE Transactions on Robotics, vol. 38, no. 5, pp. 2827–2840, 2022. doi:10.1109/TRO.2022.3164841

Full method and derivations: PhD thesis. A recorded talk on the paper is here.

What this implements

The pipeline takes an anatomical STL and returns a manipulator design — module lengths, curvature and joint parameters — optimised for that anatomy. The case study in the paper targets a knee, for arthroscopy.

• Differential Evolution (DE/rand/1/bin), written from scratch, searching a design space of roughly 10¹² candidates.
• A Monte Carlo fitness function sampling the workspace, with a two-stage forward-kinematics gate to reject infeasible candidates cheaply before the expensive evaluation.
• A voxelised task-space representation with morphologically dilated obstacles, an Obstacle Occlusion Index, and a service-sphere SO(3) coverage metric for orientation dexterity.
• HPC deployment as long-running PBS jobs with per-generation checkpointing and a resume path.

Reproducing the results

Requirements

• MATLAB with the Parallel Computing Toolbox. Developed against R2018b.
• A machine with real core count, or an HPC allocation. See the note on runtime below.

1. Anatomy scan

You need an STL of the target anatomy. The one used in the paper is a phantom knee, captured with an Artec 3D scanner — the same phantom used in the hardware experiments, so the design results and the physical validation share a geometry.

Download: kneemodel.stl (47.6 MB, units in millimetres). It is too large to sit in this repository.

Any watertight anatomical STL will work in its place; the voxelisation step below is what the optimiser actually consumes.

2. Voxelisation

GenerateVoxelisation.m

Produces a voxel map. The example output is VoxelDataMultiTarget.mat; your run will carry the date in the filename instead.

3. Optimisation

SnakeRaven_Evolution_script.m

Run it on an HPC cluster or another machine that can hold a parpool. Task objectives are switched by commenting options in and out, for example:

%Anatomyfilename = 'VoxelDataMultiTarget.mat';

This is a long-running job. The reference PBS script requests 12 parallel workers and up to 200 hours of walltime. Treat it as a cluster job, not something to run on a laptop over lunch.

The script creates a directory holding one file per fitness evaluation, named Design_alphaXXX_XXX_nXX_XX_dXXX_XXX.mat after the design parameters. Final results land in Snake_Evolution_ResultsXX-XXX_XXXX_XX_XX_XX.mat, stamped with the completion date and time.

A backup is written after every generation, so a run that dies remotely is recoverable. Point Revive_Evolution.m at the run directory and execute it to continue — also the way to extend a completed run for more generations.

Revive_Evolution.m

4. Plotting

PlotEvolutionResultsSnake_Example.m

Produces the design render, fitness over time, mean fitness and standard deviation over time, a boxplot of parameter variation, the dexterity distribution, and the maximum service sphere.

Running on HPC (PBS)

Optimisation/a_pbs_job.sh is a working example. It loads MATLAB R2018b and runs the evolution on one node with 12 processors, 5 GB of memory, and a 200-hour limit.

#!/bin/bash -l
#PBS -N SnakeRaven
#PBS -l nodes=1:ppn=12
#PBS -l mem=5gb
#PBS -l walltime=200:00:00

module load matlab/2018b

matlab -r SnakeRaven_Evolution_script -logfile logfile_SnakeRaven_Evolution_script.log -nodisplay -nodesktop -nosplash

Copy the contents of Optimisation/ to your cluster directory, then:

dos2unix a_pbs_job.sh      # if the file has been through Windows
qsub a_pbs_job.sh          # returns a job ID
qstat -USERNAME            # check status
qdel <job_id>              # abort

The logfile carries MATLAB's console output, so it is the quickest way to see which generation the run has reached.

Key files

Path	What it does
GenerateVoxelisation.m	Turns the anatomical STL into the voxelised task space.
SnakeRaven_Evolution_script.m	The optimisation entry point.
Revive_Evolution.m	Resumes an interrupted or completed run from its per-generation backup.
PlotEvolutionResultsSnake_Example.m	Regenerates the result figures.
Optimisation/a_pbs_job.sh	Reference PBS submission script.

How to cite

@article{razjigaev2022endtoend,
  author  = {Razjigaev, Andrew and Pandey, Ajay K. and Howard, David and Roberts, Jonathan and Wu, Liao},
  title   = {End-to-End Design of Bespoke, Dexterous Snake-Like Surgical Robots: A Case Study With the {RAVEN} {II}},
  journal = {IEEE Transactions on Robotics},
  volume  = {38},
  number  = {5},
  pages   = {2827--2840},
  year    = {2022},
  doi     = {10.1109/TRO.2022.3164841}
}

Related

The manipulator this designs, and the control software that drives it: SnakeRaven-Project.

Licence

MIT — see LICENSE.

Questions

Written by Andrew Razjigaev. Questions: andrew_razjigaev@outlook.com