Deep Neural Network-based Robust Collision Avoidance Control of Space Manipulator for Active Debris Removal

Tools

Sampath, Shabadini and Feng, Jinglang (2024) Deep Neural Network-based Robust Collision Avoidance Control of Space Manipulator for Active Debris Removal. In: International Astronautical Congress, 2024-10-14 - 2024-10-18, MiCo Milan Convention Centre.

[thumbnail of Sampath-IAC-2024-Deep-Neural-Network-based-Robust-Collision-Avoidance-Control-of-Space-Manipulator-for-Active-Debris-Removal]
Preview
 Text. Filename: Sampath-IAC-2024-Deep-Neural-Network-based-Robust-Collision-Avoidance-Control-of-Space-Manipulator-for-Active-Debris-Removal.pdf
Final Published Version
License: All rights reserved
Download (4MB)| Preview

Abstract

The efficiency of space debris capture relies heavily on the effectiveness of control systems deployed by space manipulators. However, it faces significant challenges due to inherent uncertainties associated with collision avoidance such as noise in the visual sensor mounted on the space manipulator, external disturbance, high-frequency noise in the control system, overshooting or oscillations in the control signals that affect the convergence and imprecise knowledge of system parameters and stability of system and inaccurately predicting trajectories. To address these challenges, a novel collision avoidance methodology integrating a Deep Neural Network (DNN) with a Proportional-Derivative (PD) controller is developed. Depending on the autonomy level, DNN is trained to autonomously initiate collision avoidance actions to improve the PD controller’s capability to control the manipulator’s motions in a complex environment by minimizing the risk of collisions. With the measurements from both the visual sensor and LiDAR, the DNN is trained to autonomously identify potential collision risks, while the PD controller controls the manipulator's motion. DNN processes the sensors' data, including visual feedback in real-time to perceive the manipulator’s environment. If the DNN detects a potential collision between the pre-planned trajectory and the target debris, the DNN overrides the control signals generated by the PD controller, plans a new trajectory using a reactive planning strategy, and controls the manipulator to avoid obstacles. Once the collision risk is mitigated, control is handed back to the PD controller, which resumes guiding the manipulator along the new trajectory toward its target position. To address the uncertainties, the DNN learns from the data from the sensors measurements and the control system’s parameters, and continuously updates its weights and biases through training, enabling it to adaptively respond to different environmental conditions and disturbances. LiDAR provides range measurements for the DNN to respond to imprecisely predicted trajectories by implementing real-time tracking and monitoring systems. The innovation of this methodology is the introduction of DNN and the seamless coordination design between the DNN-driven collision avoidance system and the PD controller, addressing uncertainties by training the DNN model to respond to any changes in the control system. Extensive simulations within MATLAB-Simulink are performed to evaluate the control performance of a DNN-based PD controller. These simulations demonstrate the controller’s effectiveness in mitigating collision risk and ensuring the safe operation of the manipulators in cluttered environments, and its robustness in overcoming uncertainties and ensuring the stability of the space manipulator for active debris removal.

ORCID iDs

Sampath, Shabadini ORCID logoORCID: https://orcid.org/0009-0004-6963-2361 and Feng, Jinglang ORCID logoORCID: https://orcid.org/0000-0003-0376-886X;
CORE (COnnecting REpositories)