Inchworm-Inspired Adaptive Multimodal Neural Control for an Autonomous Inspection Robot
Wasuthorn Ausrivong, Arthicha Srisuchinnawong, Poramate Manoonpong,∗
AI summary
Problem
Existing inspection robots struggle with complex, non-planar pipe configurations due to reliance on computationally expensive high-dimensional sensors or require manual intervention, limiting their reliability and efficiency in critical industrial environments.
Approach
The authors developed the Inchworm-inspired Adaptive Multimodal Neural Control (IAMNC), an interpretable neural framework that uses minimal exteroceptive feedback to autonomously select locomotion modes, adapt crawling gaits, and execute stable gait transitions across varying surfaces and obstacles.
Key results
- Novel IAMNC framework with three interpretable modules for autonomous multimodal locomotion and gait transitions
- General Adaptation Unit network generating primitive neural functions (reflex, memory, logic) for complex adaptive behaviors
- Real-world AVIS-II robot successfully navigates horizontal, vertical, sloped pipes, and obstacles using only nine low-cost sensors
- Demonstrated capabilities surpass prior semi-autonomous and vision-dependent inspection robots in adaptability and efficiency
Why it matters
Provides a reliable, computationally lightweight, and interpretable control strategy for industrial robots operating in unstructured or sensor-limited environments.
Abstract
While inspection robots currently achieve au- tonomous locomotion and transitions using computationally expensive multi-dimensional data, such as point-cloud maps, in- vertebrates like inchworms can navigate complex tree branches effortlessly with a simple brain and little information. Inspired by this, an Inchworm-inspired Adaptive Multimodal Neural Control (IAMNC) is presented here. It relies on only nine exteroceptive sensors to achieve autonomous locomotion and gait transitions of a bipedal out-pipe inspection robot. The IAMNC consists of three interpretable modules: Mode Selection (MoSe) for autonomous mode selection, Locomotion Control (LoCo) for adaptive crawling, and Transition Control (TraCo) for adaptive transitions. Additionally, this work also proposes interpretable Adaptation Units (AUs, part of TraCo), which can be configured to obtain different functions (e.g., reflex, memory, and logic gate operations). Thus, connecting them as an Adapta- tion Unit Network (AUN) results in complex yet understandable adaptation signals for autonomous and stable gait transitions. With this control approach, the robot demonstrates autonomous locomotion, gait transitions, and adaptability across various pipe connection types. It can also autonomously step over an obstacle on the pipe, relying on seven infrared obstacle detection sensors (IR sensors) and two inductive sensors, rather than a computationally expensive camera sensor.