Aaron Tran

Aaron Tran

Control Software Engineer at Brembo Coding Hub

Robotic Arm Optical Coherence Tomography Motion Planning

This project is aimed at enabling the use of optical coherence tomography with bedridden patients in potentially cluttered environments, as one might find in a hospital room. My contributions toward this are in the path planning of a UR5e robotic arm, subject to the following constraints:
  1. The patient must remain in view through motion
  2. The robot must respect prescribed kinodynamic limits to ensure patient safety
  3. Generated motion plans must be both smooth and predictable, as to inspire confidence in the patient
  4. The plan must be capable of replanning at 10Hz to support a dynamic environment
Sample based planners were written due to their prevalence over the past few decades. Probabilistic trajectory optimization is a more modern approach, that was investigated towards the end of this project.

Sampled Based Planning

Sample based planners were initially written in MoveIt2 and moved to Klamp't for easier prototyping and greater control over constrained sampling. Single-Query Bi-Directional Probablistic Roadmap with a shortcut heuristic was selected as the planner of choice due to its ability to quickly generate plans. The constraint sampler allows for sampling poses in task space that force the robot to stay on the same side of singularities (elbow up/down, wrist write/left, etc.) and keeping the target eye within view .

Obstacle Free


Obstacle Present


Note that the jerkiness of this demonstration is due to passing an unprocessed plan through a trajectory optimizer from Klamp't

Sample Based Trajectory Optimization

This method for path planning has seen considerable attention from research in recent years. It is also capable of determining direct torque inputs to obtain prescribed control goals, but for our case just determining a path is sufficient. In comparison to traditional sample based planning, this method can produce more rich behabior through careful cost function design. They also naturally support quick replanning, as they work by only rolling out the first action from each optimization.
The two algorithms considered are model path predictive integral control (MPPI) and improved cross entropy method (iCEM). They were developed using Stochastic Tensor Optimization for Robot Motion, a toolkit designed specifically for designing algorithms such as these, which rollout many particles in parallel.

MPPI


iCEM