An intelligent spinal soft robot with self-sensing adaptability

Abstract

Self-sensing adaptability is a high-level intelligence in living creatures and is highly desired for their biomimetic soft robots for efficient interaction with the surroundings. Self-sensing adaptability can be achieved in soft robots by the integration of sensors and actuators. However, current strategies simply assemble discrete sensors and actuators into one robotic system and, thus, dilute their synergistic and complementary connections, causing low-level adaptability and poor decision-making capability. Here, inspired by vertebrate animals supported by highly evolved backbones, we propose a concept of a bionic spine that integrates sensing and actuation into one shared body based on the reversible piezoelectric effect and a decoupling mechanism to extract the environmental feedback. We demonstrate that the soft robots equipped with the bionic spines feature locomotion speed improvements between 39.5% and 80% for various environmental terrains. More importantly, it can also enable the robots to accurately recognize and actively adapt to changing environments with obstacle avoidance capability by learning-based gait adjustments. We envision that the proposed bionic spine could serve as a building block for locomotive soft robots toward more intelligent machine-environment interactions in the future.

Graphical abstract

Figure 1

Depiction of high-level intelligence in both living creatures and soft robots (A) Spines for self-sensing, regulation, and environmental adaption. Vertebrate animals rely on spines for sensing and supporting and muscles for actuation. Proposed soft robots rely on bionic spines for sensing, support, and auxiliary actuation and origami muscles for actuation. (B) Demonstration of an autonomous self-sensing adaption process and methodology on the decoupling mechanism of a unibody sensing-actuation-integrated spine. Scale bar: 5 cm.

Figure 2

Actuation mechanism and characterization results of the spine-assisted actuation (A) Illustration of the actuation mechanism. (i) Contraction of the artificial muscle under vacuum pressure; (ii) recovering under the elastic force, release force, and auxiliary force generated by the artificial muscle and bionic spine. (B) Normalized recovery force under auxiliary actuation as a function of actuation voltage and actuation frequency, i.e., 1 Hz, 5 Hz, and 20 Hz from top to bottom, with the pneumatic actuation at 1 Hz. (C) Measured stride length of the crawling robot without terrain contact as a function of auxiliary frequency for three different pneumatic actuation frequencies, i.e., 1 Hz, 2 Hz, and 3 Hz. (D) Measured stride length of the robot as a function of actuation frequency with and without auxiliary actuation while the auxiliary voltage is 1 kV. (E) Comparison of robot crawling speed with and without auxiliary actuation on the rubber land. (F) Measured crawling speed of the robot at an actuation frequency of 1 Hz for different terrains. (G) Measured crawling speed of the robot as a function of actuation frequency on the rubber land.

Figure 3

Environmental awareness and recognition by the spinal robot (A) Measured feedback signal by the spine when the robot moves with an increased actuation frequency from 1 Hz to 20 Hz. (B) Measured feedback signal of the spine when the robot crawls on different terrains as a function of actuation frequency. (C) Images of the crawling robot on different terrains, including some relatively smooth terrains, i.e., rubber land, PET land, acrylic land, and plush land, and some uneven terrains, i.e., grass land, gravel land, and stone land. (D) Flow diagram of the machine learning process. (E) Classification confusion matrix of the robotic recognition on single-terrain tasks, with an overall accuracy of 98%. (F) Classification confusion matrix of the robotic recognition on multi-terrain tasks, with an overall accuracy of 97.1%.

Figure 4

Self-sensing adaption by the spinal robot in multi-terrain tasks (A and C) The unimodal crawling robot moves over two multi-terrains scenarios, as the first scenario (A) contains rubber land, stone land, and PET land, and the second scenario (C) contains grass land, plush land, and gravel land. (B) The spinal crawling robot with self-sensing adaption for the first scenario, overcoming the terrain boundary that the unimodal one failed. (D) The spinal crawling robot for the second scenario, showing it to be 18% faster than the unimodal one. (E) Self-sensing adaption of the spinal robot remains available in dusky and dark environments. (F) Images of the spinal crawling robot performing a self-adaptive multi-terrain-crossing task in dazzling light condition. (G) Images of the spinal crawling robot performing another self-adaptive crossing task in purely dark condition.

Figure 5

An amphibious and omnidirectional spinal soft robot (A) Design and reversible locomotive mechanism of the robot consisting of twin actuators in parallel. (B) Omnidirectional motions of the robot. (C) Images of the robot performing a self-adaptive obstacle avoidance with self-sensing adaption and dexterous motion. (D) Images of the robot performing multi-terrain crossing and amphibious transition with self-sensing adaption.