Occupancy-Aware Neural Distance Perception for Manipulator Obstacle Avoidance in the Tokamak Vacuum Vessel.

Area of Science:

• Robotics
• Computer Vision
• Geometric Deep Learning

Background:

• Accurate distance perception and collision reasoning are vital for robotic manipulation within confined spaces like tokamak vacuum vessels.
• Traditional mesh- or voxel-based methods present limitations including discretization artifacts, discontinuities, and high memory usage, hindering continuous geometric reasoning and optimization-based planning.

Purpose of the Study:

• To present a novel Occupancy-Aware Neural Distance Perception (ONDP) framework as a compact, differentiable geometric sensor for manipulator obstacle avoidance in reactor-like environments.
• To introduce a Physically-Stratified Sampling strategy to address sampling inadequacies in constrained environments by basing data distribution on engineering constraints.

Main Methods:

• Developed an Occupancy-Aware Neural Distance Perception (ONDP) framework utilizing a lightweight neural network trained in physical units (millimeters).
• Implemented a Physically-Stratified Sampling strategy with weighted quotas and symmetric boundary constraints for robust gradient learning in critical safety regions.
• Employed a mean absolute error loss function to ensure strict adherence to engineering tolerances.

Main Results:

• Achieved approximately 2-3 mm near-surface accuracy with high-frequency distance and normal queries (exceeding 15 kHz for large batches, a 5911x speed-up over mesh-based queries).
• Demonstrated continuous, sub-centimeter geometric fidelity in experiments on a tokamak vessel model.
• Validated the ONDP framework's capability for real-time perception, monitoring, and motion planning.

Conclusions:

• The ONDP framework provides a breakthrough in geometric sensing for confined-space robotic manipulation.
• Its high accuracy, speed, and continuous geometric fidelity enable seamless integration with trajectory optimization and model-predictive control.
• This advancement significantly enhances robotic capabilities in challenging environments like tokamak interiors.