The number of soft robotics publications in 2010 and 2011 average 340 per year and are fairly close, as shown in Figure 2. Then, the yearly publications increase significantly by 38% to 459 in 2012. The increasing trend continues with yearly increments of publications from dozens to a few hundreds during 2012–2015. Starting 2016, the rising rate of yearly publications has been relatively stable, with high percentages of 33%, 24%, 25%, and 22% in 2016, 2017, 2018, and 2019, respectively. It is worth noting that the yearly publications are more than 1,000 for the first time in 2017. The active interests and intensive efforts from robotics community in the past few years has led to the significant progress of this field, which is indicated by the enormous increase in the number of publications. Since 2020, the yearly publications increase with a lower growth rate. Compared to 1715 publications in 2020, 900 publications have been contributed in the first 7 months this year, indicating that the growth has started to slow down.

In total, 101 countries or regions contribute to our final publication collection. Table 1 lists the top 20 most productive countries or regions in soft robotics field ranked by the number of total articles (TA) during the last 12 years. It is observed that China is the most productive country with a total publication of 2966, followed closely by the United States (2911), then Japan (1048). However, in terms of average citations per publication (ACPP), the top three countries/regions are the United States (28.82), the Netherlands (28.55), and Belgium (26.36), indicating stronger impact per publication. It is worth noting that, despite being the most productive country, China’s ACPP is relatively low. The situation is similar for Japan. It is unclear whether it is due to research quality itself, language bias, or joining this field late. Another observation is that a large number of publications are international-collaborated articles (28.2%) for the top 20 country/region’s publications, indicating soft robotics has attracted worldwide researchers to work together. In addition, the top four countries also have the largest number of collaborative countries. It is also noted that ACPP is generally higher for international articles, implying that collaborations may have facilitated higher-quality research, hence greater impacts. It may also explain why the Netherlands and Belgium have relatively high ACPP, as their international-collaborated articles account for 65% and 61% of their total publications.

Table 2 lists the most productive 20 institutions in the soft robotics field. Apparently, most of them are from the top 10 most productive countries and China and the United States account for seven and four institutions, respectively. Chinese Academy of Sciences leads the list with the most publications followed by University of California System and Harvard University. In terms of ACPP, institutions from the United States lead the list with Harvard University (74.14) and Massachusetts Institute of Technology (59.81). Apparently these top institutions have taken a prominent role in developing and promoting soft robotics. Except for institutions in the United States, Scuola Superiore Sant Anna (36.09), University of Wollongong (36.09), and Seoul National University (27.96) have the highest ACPP.

Soft robotics is a interdisciplinary field, where researchers from a number of subject areas have contributed to develop this field. The related subject areas and their proportional contribution in the field are depicted in Figure 3. The areas are assigned by WoS. It is observed that Engineering is the most prominent subject area for soft robotics research, followed by Chemistry, Others, and Computer Science. To be more specific, Robotics (19.13%), Engineering Electrical Electronic (10.86%), and Automation Control Systems (7.19%) are ranked as the top three in the Engineering area. On the other hand, Nanoscience Nanotechnology (6.34%), Chemistry Physical (4.48%), and Polymer Science (4.24%) account for the top three in the Chemistry area. Computer Science Artificial Intelligence (5.81%) is the top one sub-area in the Computer Science area. Futhermore, Others include Physics, Instruments Instrumentation, Medical Science, Neurosciences, Energy Fuels, etc., which suggests that the soft robotics is a highly diverse field involving interdisciplinary subject areas. Based on the main themes of these subject areas, several related research topics can be identified: (1) electronics and control technologies for actuation, sensing, and control, (2) advanced materials and their applications in soft robotics, and (3) artificial intelligence enabling intelligence for soft robotics.

There are a number of journals associated with soft robotics due to the inherent interdisciplinary subject areas involved in soft robotics. Table 3 lists the top 20 journals ranked by the total number of articles and their related impact indicators. It is observed that the majority of the journal’s categories belong to the subject areas and sub-areas in Section 3.4. IEEE Robotics and Automation Letters has published the most articles (377), followed by Soft Robotics (288) and Smart Materials and Structures (249). In terms of impact factor (IF), Advanced Materials (30.254) has the highest score among all journals and Soft Robotics (9.904) leads the list in the Robotics category. Regarding average journal impact factor percentile (AJP), Advanced Materials (97.578%), Nature Communications (95.205%), and Soft Robotics (94.643%) rank as the top three journals. Apart from robotics-oriented journals, the top journals contributing to soft robotics are in the fields of Chemistry and Material Science such as Advanced Materials, Advanced Functional Materials, Soft Matter, Smart Materials and Structures, ACS Applied Materials Interfaces etc., suggesting materials research play a key role in the soft robotics field.

Each research paper usually cites a number of references. In a co-citation network, these references are represented as nodes. The degree of connectivity between the nodes of these references indicates how frequently they are cited by the same articles. The assumption is that if two references are frequently cited together, the two references must be related in some way. The inherent connectivity reflects the comprehensive relationship between the cited references and the articles that cite them.

The overview of the co-cited network on soft robotics depicted with CiteSpace is shown in Figure 4. Each node is a reference cited by publications in our WoS collection and is depicted with a series of citation tree-rings across the series of time slices. Note that the nodes are not necessarily in our WoS collection, but the references cited by our WoS collection. The size of a node means the number of citations the associated reference has received: the larger a node, the more citations the node has received. The color of a node’s rings correspond to the time it is cited: the darker blue, the earlier it was cited; and the more red, the more recent it was cited. The thickness of the red rings outside of many nodes indicates how many recent citations that reference has received: the thicker the ring, the more citations; and the more red the ring, the more recent the citations. A red dot at the center of a node corresponds to citation bursts, meaning sudden increases of citations are detected. The color of nodes can be used to trace the timeline of citations and citation bursts are useful to trace the emerging research topics. The connecting curves indicates how often the references are cited together: the thicker the curves, the more often the references are co-cited.

CiteSpace generates 19 clusters based on the association between the nodes, and the areas of different clusters are filled with different colors, as shown in Figure 4. Again, blue and red areas mean earlier and more recent citations, respectively. The clusters are named from titles of citing articles of the cluster using CiteSpace’s built-in algorithm LLR (log-likelihood ratio). Similar to the co-occurrence map, the LLR algorithm also computes based on the number of co-occurring events and is widely used in text computing (Kim and Chen, 2015). The clusters’ ID numbers are sorted by the number of nodes inside: the more nodes inside that cluster, the smaller the ID number. For example, the cluster variable stiffness has the most nodes and is labeled as 0 by CiteSpace.

Table 4 lists the top 10 clusters by their size, that is the number of nodes or members in each cluster. Clusters with few members tend to be less representative than larger clusters because small clusters are likely to be formed by the citing behavior of a small number of publications. The quality of a cluster is also evaluated by the silhouette score (S-score), which represents the cluster’s homogeneity or consistency. Homogeneous clusters tend to have silhouette values close to 1. The silhouette scores of the clusters range from 0.817 (Cluster #2) to 0.987 (Cluster #9), suggesting that our clustered members are highly homogeneous. The average years of publication of the clusters ( Y ̄ ) indicates the cluster’s recentness. For example, Cluster #3 has an average publication year of 2008. The most recently formed cluster in Table 4, Cluster #6 on liquid crystal elastomer, has an average year of 2017.

The most cited 10 references by our WoS collection are listed in Table 5 and labeled in black in Figure 4, and the larger the font, the more citations that reference has received. They are considered as landmark articles due to their groundbreaking contributions. Cluster #0 has 8 articles in the top 10 landmark articles, four of which are general reviews of the soft robotics field at that time. Each of Clusters #5 and #10 has one. The most cited article in our dataset is a review article by Rus and Tolley (2015) with 1052 citations. The second one is by Wehner et al. (2016) with 372 citations. They reported a pure soft and untethered robot composed of on-board fuel reservoirs and catalytic reaction chambers. Articles at the third (Kim et al., 2013), fifth (Laschi et al., 2016), and seventh (Majidi, 2014) are all reviews from Cluster #0. The article at the fourth is by Mosadegh et al. (2014), which described a new design for a pneumatic network (pneu-nets) that increases the actuation speed significantly. The sixth article reports a soft robotic glove for combined assistance and at-home rehabilitation (Polygerinos et al., 2015). The article at the eighth presents an untethered soft robot made of pneu-nets (Tolley et al., 2014). The ninth article from Cluster #10 demonstrated artificial muscles made of fishing line and sewing thread (Haines et al., 2014). The article at the 10th is a review from Cluster #5.

Figure 5 visualizes the timeline of distinct co-citation clusters, which is a presentation of Figure 4 by year. It is observed that the highly-cited references in Clusters #3, #4, #8, #9, #11, #13, #14, #16, #19 were mainly published before 2015, meaning they were hot topics earlier. Clusters #15 and #17 spanned during 2017–2019 and 2015–2019, respectively, suggesting they were hot topics a while ago but lacking of evidence for the latest impactful research. Another observation is that Clusters #0, #1, #2, #5, #6, #7, and #10 have been published and cited during the past decade, indicating continuous research interests, and we will focus on analyzing these clusters in the top ten largest clusters, i.e., excluding Cluster #10.

Table 6 lists the most cited 5 references for each of Clusters #0, #1, #2, #5, #6, and #7. Here, cites are the number of citations cited by our WoS collection. It is observed that 1/3 of the highly-cited references are review articles (indicated by a), which is also a sign that the knowledge of soft robotics has been advancing rapidly.

Cluster #0 on variable stiffness includes numerous nodes with red rings of citation bursts indicating a continuous hot topic. Variable stiffness enables high performance and safe human-robot interaction for soft robotics simultaneously. A review by Rus et al. (Rus and Tolley, 2015) is the most cited article both in Cluster #0 and by our WoS collection. This article examines the then developments of soft robotics and major design components, and proposes ideas for the future development including programmable stiffness, proprioceptive sensing, contact modelling etc., which had a profound impact on the soft robotics community. The two regular papers (Mosadegh et al., 2014; Wehner et al., 2016) except for the two review papers (Kim et al., 2013; Laschi et al., 2016) in Clusters #0 focus on pure soft robots and fast soft actuators, respectively. The top cited references in Cluster #0 seems to be not specifically on variable stiffness. It is unclear whether it is due to the limitation of using terms from citing articles to name clusters, or some of them mentioned variable stiffness as an important attribute which inspired roboticists to develop variable-stiffness soft robotics.

The core members of Cluster #1 represent major milestones in relation to strain sensor, notably Zhao et al.’ article on stretchable optical waveguides for strain sensing in a prosthetic hand (Zhao et al., 2016b), Larson et al.’s article on stretchable electroluminescent skin (Larson et al., 2016), and Muth et al.’s article on 3D Printing of strain sensors (Muth et al., 2014). In addition, Amjadi et al. (2016) and Dickey (Dickey, 2017) reviewed stretchable sensors and soft-and-stretchable electronics using liquid metal, respectively. The above indicate a trend for the application and manufacturing of stretchable sensors, which enable potential for exteroception, proprioception, and wearables soft robotics.

Cluster #2 contains further milestones in multiresponsive actuator. Hydrogels, as a sensitive type of polymer capable of producing large, reversible deformations under multiple external stimuli such as temperature, light, and pH, have attracted great attention in the field of soft actuators. The most cited work in this cluster is the 2017 article by Yuk et al. (2017). Their work pioneered the high-speed, high-force, and optically and sonically camouflaged hydrogel actuators. Another prominent member of the cluster is the 2015 article by Breger et al. (2015). Their work reported self-folding thermo-magnetically responsive soft microgrippers and opened up new grounds for multiresponsive soft robots. Morales et al. created the micro tool clamp by coupling photo-cross-linked pNIPAM-AAc soft-hydrogel with a nonswellable and stiff segmented polymer, which relies on water penetration to provide a swelling response to external stimuli (Morales et al., 2014). Other materials have also been explored in addition to hydrogels. Taccola et al. (2015) developed soft actuators controlled by either joule effect or environmental humidity variation using PEDOT:PSS and PDMS. These research also reflect that chemistry and material science play vital roles in soft robotics, which is consistent with our previous findings.

Another observation is that the mean years of the largest 3 Clusters #0, #1, and #2 are around 2015. In addition, around this time, a large number of researchers systematically reviewed the whole field of soft robotics and conducted research from various aspects, which may have helped to define multiple key aspects of soft robotics and accelerate the development of soft robotics. These suggest that these clusters are structurally essential and inspirational. The above may also explain why the soft robotics field started to take off since 2016 in Figure 2.

Cluster #5 on 4D printing, with an average year of 2016, is one of the youngest clusters. 4D printing is a manufacturing method that combines stimuli-responsive matter with 3D printing to create robots that can self-transform into specified shapes in response to external excitation. This technique eliminates the need for molding which was the previous major manufacturing approach and enables rapid prototyping of soft robots. The most cited article in this cluster, Gladman et al. (2016), demonstrated how bionic hydrogel composite can be directly manufactured using 4D ink-jet printing technology. Their work is a major milestone. Another prominent work, Kim et al. (2019), reported direct ink writing of elastomer composites containing ferromagnetic microparticles and suggested new possibilities for applications in flexible electronics, biomedical devices and soft robotics. It is noted that this cluster also includes the publication with the highest citation burst in 2020, which will be further analyzed in Section 5.

Cluster #6 on liquid crystal elastomer (LCEs) is the most recently (2017) formed cluster. The inherent mechanooptic and thermomechanical effects in LCEs allow for both light and heat actuation, enabling multiresponsive actuators and remote powering and control for untethered soft robotics. The one that has the highest citation is a review article by White et al., in 2015 (White and Broer, 2015). It reviewed liquid crystalline polymeric materials with emphasis on the thermally and photogenerated macroscale mechanical responses, which opened up opportunities for mechanical designers to employ these materials in soft robotics. The second most cited work is by Palagi et al. (2016) and they developed a robot with LCEs that can be propelled by structured monochromatic light. In addition, the rest of the 3 highly cited articles all present efficient microactuators made of LCEs (Ware et al., 2015; Gelebart et al., 2017; Wani et al., 2017).

Cluster #7 is also a relatively young cluster with an average year of 2016, indicating that the dielectric elastomer actuator (DEAs) is likely an emerging research topic. DEAs are composed of a dielectric elastomeric film sandwiched between two compliant electrodes. By applying a high voltage across the electrodes, which generates an electrostatic attraction between them, squeezing the elastomer membrane, resulting in elastomer thickness reduction and area expansion. DEAs stands out because of their exceptionally fast response, light weight, and large actuation stroke. The most cited paper in this cluster, Shintake et al. (2016), presented a novel electrode arrangement electrode arrangement for soft grippers to optimize electroadhesion and electrostatic actuation simultaneously, generating over ten times higher electroadhesion force than would be obtained from a conventional DEA electrode configuration. The second most cited article is by Li et al. (2017). They developed a fast-moving soft electronic fish by combining DEAs and ionically conductive hydrogel, resulting in both high speed and consistent performance. Another article by Acome et al. (2018) addressed the problem that traditional DEAs actuators are likely to fail due to electrode breakdown or electrical aging by using self-healing liquid dielectric. The second most cited reference in this cluster is a review article by Rich et al. (2018), where they pointed out that improving DEAs’ output force and stroke will enable further applications in untethered systems. For a detailed review of DEAs actuated soft actuators, readers are encouraged to refer to (Gu et al., 2017).

It is noticed that some of the clusters have similar names but form different clusters, such as clusters #6 (liquid crystal elastomer) and #11 (crosslinked liquid-crystalline polymer). We further investigate the differences between these clusters. The references in cluster #11 focus on the early pioneering studies of applications of liquid crystal elastomers, such as Van Oosten et al. (2009) and Ikeda et al. (2007). Cluster #6 includes the widespread use of liquid crystal elastomers for soft robotics in more recent times. This difference can also be noted from Figure 5. Cluster #9 (pneumatic artificial muscle) includes many references related to pneumatic muscles (e.g., Lin et al. (2015); Sárosi et al. (2015)), while quite a few references in cluster #16 (pneumatic muscles-a case study) involve commercialized robotics, such as Albu-Schäffer et al. (2007a), Albu-Schäffer et al. (2007b)) and it is a relatively older cluster. Cluster #10 (artificial muscle) is concerned with general artificial muscles. A representative article uses polymer fibers, such as fishing line, to create artificial muscles (Haines et al., 2014). Since the clusters are formed by co-citation behavior, it is expected that the clusters can be formed by co-citation at different times, different actuation methods, etc., which differentiates them from similar clusters. In this way, these clusters may include different information on similar fields and trends at different times.

Citation bursts are useful to trace research focus. A citation burst has two key aspects: the strength and the duration of the burst. Table 7 lists the top 20 references with the strongest citation bursts across the entire dataset during the period of 2010–2021. It is observed that 19 of the 20 references were published between 2008 and 2014, with the majority of their citation bursts occurring between 2010 and 2018, indicating that these references are pioneering work that had significant impacts on the soft robotics community, especially during the early days. Among the top 20 references with strongest citation bursts, Shepherd et al. (2011), titled “Multigait soft robot,” is on the top of the list with the burst strength of 91.96. It has the strongest citation burst, which started in 2013, the third year after publication, and ended in 2016. This article published in Proceedings of the national academy of sciences proposed the concept of a locomotive quadrupedal soft robot. The second article (Kim et al., 2013) with strongest citation bursts is one of the earliest reviews on soft robotics. The next article is by Ilievski et al. (2011), the author who used pneumatic networks (PneuNets) to create large-amplitude actuations in soft elastomers by pressurizing embedded channels. Notably, almost half of the references focus on various soft actuators (Brochu and Pei, 2010; Ilievski et al., 2011; Anderson et al., 2012; Lima et al., 2012; Martinez et al., 2013; Haines et al., 2014; Mosadegh et al., 2014) and manipulators (Laschi et al., 2012; Martinez et al., 2012). The above suggest that soft actuators have been the core aspect of the soft robotics research. In addition, 5 references were on locomotive robots with focus on locomotion (Lin et al., 2011; Shepherd et al., 2011; Seok et al., 2013; Tolley et al., 2014) and camouflage (Morin et al., 2012), indicating the interests of exploring unpredictable environments with soft robots. The most recent reference with the strongest citation bursts is by Kim et al. (2018), which reported 3D printing of programmed ferromagnetic domains in soft materials that enable fast transformations between complex 3D shapes via magnetic actuation. This work represents two of the emerging trends in soft robotics, ferromagnetic composites and 4D printing, which will be discussed further in Section 5.

New materials enable new actuation, sensing, control, and applications, as well as require new integration methods, which expands the soft robotics field as a whole. It was found previously that materials related research were critical in soft robotics field: in Section 3.4, Chemistry was found to be the second largest research subarea in soft robotics; and multiple clusters were named after certain materials and materials integration method including Cluster #3, #4, #5, #6, #7, #8, and #11. More importantly, multiple clues have led us to believe that new materials and materials integration are of the most recent emerging trends and potential future directions in soft robotics: the youngest clusters are named after certain materials and integration (Cluster #5, #6, and #7); and the vast majority of the articles in Table 8, which includes both references with recent citation bursts and citing articles with high cites, are related to the development of new materials and composites, as well as materials integration methods.

Multiple materials and composites are gaining momentum. Soft magnetic composites seem to attract the most attention as they have can be controlled by remotely controllable actuating fields and high penetration depths, enabling wireless applications such as drug delivery. These composites are made with various soft polymers embedded with uniformly dispersed magnetized or magnetizable microparticles such as neodymium-iron-boron (NdFeB) particles. These composites often appear as mobile microrobots and are referred to as ferromagnetic soft robots by Kim et al. (2019). Hu et al. (2018) developed a magneto-elastic soft millimetre-scale robot using silicone elastomer embedded with hard NdFeB microparticles, which showed high mobility through multimodal locomotion. Kim et al. (2019) presented self-lubricating soft continuum robots with soft polymer matrix composed of polydimethylsiloxane (PDMS) or thermoplastic polyurethane (TPU), hard NdFeB particles, and a hydrogel skin. It was demonstrated that the robot was able to achieve omnidirectional steering, navigate in a vascular phantom, and deliver laser. Multiple other research efforts have attracted recent attentions by enabling magnetically driven robots with various functional features such as camouflage (Du et al., 2020), sensing and wireless communication (Lu et al., 2020), and shape memory (Ze et al., 2019). A particular interesting work by Li et al. (2020) manipulated droplets by a magnetic-actuated robot made of two steel beads, resulting in a system with blurry boundaries between rigid, fluidic, and soft robots.

Liquid crystal elastomers (LCEs) as a thermally driven actuating material that combines polymer network and liquid crystal mesogens, have also received significant citation bursts. LCEs can be actuated by multiple actuation sources such as direct environmental heating, light through photothermal photochemical effects, or electrical power via combination with resistive heaters, enabling multiresponsive actuators. What many of these recent LCEs related articles have in common appear to be the focus on proposing various techniques of manufacturing and exploring multimodal locomotion. Zhang et al. developed a manufacturing approach for simultaneous welding and aligning LCE materials with different chemical compositions and physical properties by thermal polymerizing prefabricated LCE films with reactive acrylate groups (Zhang et al., 2020). Using this method, multicomponent/multimaterial three-dimensional (3D) LCE robots with multiple 3D geometries were created. Saed et al. (2018) used DIW (direct ink writing) to create molecularly-engineered liquid crystal elastomer actuators. He et al. (2019) created an LCE tubular actuator by sandwiching loosely cross-linked LCE with heating wires, which was then UV irradiated to form a tubular shape. It was demonstrated that by powering the heating wires inside it, multi-directional bending and uniform contraction can be achieved. Xiao et al. reported a soft robots capable of executing various types of biomimetic locomotion, which were also actuated electrically by heating wires (He et al., 2019).

Hydrogels and dielectric elastomers have also gained increased citations. Hydrogels can provide space for ion diffusion transfer and have demonstrated many advantages including ionic conductivity, stretchability, biocompatibility, optical transparency, self-healing capability etc., which makes them good candidates for flexible sensors, conductors, and actuators. Liu et al. designed an anti-freezing hydrogel with zwitterionic poly (ionic liquid) (PIL) for use in a multimodal artificial skin (Liu et al., 2020), which exhibited super stretchability (about 900%), self-healing ability, and high electrical conductivity. Yang et el. developed stretchable and transparent double-network hydrogel ionic conductors which can be used for both sensing and conducting in soft systems (Yang and Yuan, 2019). Mishra et al. reported a soft hydrogel-based actuator that can maintain stable body temperatures via autonomic perspiration (Mishra et al., 2020). Dielectric elastomers have been used for fast actuators and stretchable sensors during the past decade. One particular interesting work by Tan et al. developed a self-healable, low-field illuminating optoelectronic stretchable device by introducing a transparent, high permittivity polymeric dielectric material (Tan et al., 2019).

New materials need new materials integration methods. Direct printing attracted the most attention as 13 (24%) articles (Gao et al., 2018; Hong et al., 2015; Huang et al., 2020; Kim et al., 2018; Liu et al., 2018; Saed et al., 2018; Skylar-Scott et al., 2019; Truby et al., 2018; Wallin et al., 2018; Yuk et al., 2020; Zhao et al., 2020; Mishra et al., 2020; Zhang et al., 2020) in Table 8 focus on 3D or 4D printing soft robots, which cover a variety of topics including biomedical devices, drug delivery micro-robots, reconfigurable robots, flexible electronics etc. So-called 4D printing (Tibbits, 2014) combines 3D printing with programmable and stimuli-responsive materials, where the fourth dimension corresponds to a later shape change from the printed structure excited by various external stimuli (heat, electricity, light etc.). The majority of the above publications still use the term 3D printing, indicating the term of 4D printing is not fully differentiated from 3D printing yet. However, the directly printed stimuli-responsive soft robots which later could be controlled by external stimuli already differentiated them from traditional 3D printing as a fact. In the rest of this paper, we will use 3D printing, 4D printing, or direct printing alternatively. In 2018, Wallin et al. (2018) reviewed various 3D-printing materials and printing methods of soft robots such as 3D printing, direct ink writing, fused deposition modelling, inkjet printing, and stereolithography.

3D printing of soft robots provide an excellent, systematic tool to directly fabricate soft robots, which enabled the design of soft robots with sophisticated motion capabilities (jumping, complex 3D movements, gripping and releasing etc.) and integrated sensing. Corresponding to the materials discussed in Section 5.1, 3D printing is used for creating soft robots with ferromagnetic composites (Kim et al., 2018), LCEs (Saed et al., 2018; Zhang et al., 2020), and hydrogels (Hong et al., 2015; Gao et al., 2018; Huang et al., 2020; Mishra et al., 2020). Different materials require different printing methods. Several other materials have also attracted much attention given the variety of materials used in soft robotics. Yuk et al. (2020) 3D printed conducting polymer based on PEDOT:PSS, and further fabricated a soft neural probe capable of in vivo single-unit recording. Zhao et al. (2020) developed a direct ink writing protocol to create miniaturized silica aerogel objects. A particular interesting work by Liu et al. (2018) developed a new method and material system capable of 3D-printing living responsive materials and devices. Furthermore, multiple nozzles can be used to deposit different polymers or inks, leading to multi-material prints. For example, a millipede-like soft robot that locomotes by co-printing multiple epoxy and silicone elastomers (Skylar-Scott et al., 2019), and fabrication of soft somatosensitive actuator innervated with multiple soft sensors (Truby et al., 2018). In summary, a series of recent articles have developed various 3D printing systems for a variety of stimuli-responsive soft materials, and multi-material printing have showed the potential for direct printing of soft robots with complex architectures, multimodal motion, and embedded sensing.

Soft systems, in contrast to conventional stiff robots, and other piecewise rigid systems with finite degrees of freedom, are continuous elastic bodies with infinite degrees of freedom. This poses new challenges in sensing and proprioception, feedback and adaptive control, path planning, and robot intelligence, all of which are beyond the scope of conventional robotic algorithms. Furthermore, the development of new materials, actuators, and integration methods will require new sensing and intelligence methods in the mean time. Multiple evidences have indicated the emerging trends for sensing and intelligence: two clusters are directly named after sensing including Cluster #1 and #15 on strain sensor and control sensing, respectively, where Cluster #1 is also the second largest cluster and a relatively young one with the average year of 2015. In addition to multiple publications related to sensing discussed previously, 15 articles (about 28%) in Table 8 focus directly on sensing and intelligence of soft robots.

Multimodal stretchable sensing skins have attracted significant attention lately, which are mainly composed of various highly-stretchable conductive polymers and structures. An article by Kim et al. (2014) received widespread attention and citation bursts during 2018–2019. They developed a highly integrated stretchable artificial skin using ultra-thin single crystal silicon nanoribbons (SiNR) sensor arrays and humidity sensors capable of multimodal sensing including strain, press, temperature, and humidity. External stimuli were demonstrated to transfer via prosthesis to specific nerves in human bodies, mimicking real skins. Lei et al. fabricated a supramolecular biomimetic skin using supramolecular polyelectrolyte hydrogels capable of detecting temperature, strain, and stress (Lei and Wu, 2018). Hua et al. created a highly stretchable and adhesive matrix network (SCMN) artificial skin that can detect temperature, intra-surface strain, relative humidity, UV rays, magnetic fields, pressures, and proximity, enhancing multimodal sensing in significantly (Hua et al., 2017). Wang et al. (2019) fabricated a self-healable ternary polymer composite to detect strain and pressure. Han et al. (2020) developed catheter-integrated soft multilayer electronic arrays for multiplexed sensing and actuation during cardiac surgery, which can detect temperature, pressure and electrophysiological parameters.

Multiple other sensors have received increasing attention as well. Li et al. proposed a superhydrophobic paper-based strain sensor which could detect ultra-low strain as low as 0.1% and showed potential for application in wearable devices (Li et al., 2019). Wang et al. developed a compliant ultrathin sensing and actuating electronics innervated fully soft robots with sensors made of single-crystal Si optoelectronic photodetectors, which can sense the environment and perform soft bodied crawling adaptively (Wang C. et al., 2018). Jin et al. demonstrated triboelectric nanogenerator sensors to capture the continuous motion and tactile information for soft gripper (Jin et al., 2020).

It is observed an emerging trend of combining sensing data with data-driven learning algorithms to process complex information such as robots’ own status like shape, and interface information like force. Some of these sensors are even commercially available ones. Homberg et al. (2015) were one of the earliest to adopt learning algorithms in haptic identification of objects with a soft gripper, where commercial resistive flex sensors were used. Elgeneidy et al. (2017) derived both neural network and regression models based on experimental sensing data, and the resulting empirical model was effectively used in closed-loop control. Thuruthel et al. (2019) developed a long short-term memory (LSTM) network with redundant and unstructured sensor topology embedded in soft actuators, which was used for kinematics and force modeling of soft actuators in real time. Zhu et al. (2020) created a tactile feedback smart glove with a triboelectric tactile sensor, and trained a convolutional neural network (CNN) model for object recognition. Meerbeek et al. (2018) reported soft optoelectronic sensory foams with proprioception using multiple machine learning algorithms, where the magnitude and type of deformation for soft actuators can be identified.

In summary, the sensing modality is expanding and approaching human sensations including pressure, vibration, pain, humidity, temperature, etc., which are joint sensing results of various stimuli. Soft roboticists are trying to detect various stimuli accurately, improve sensitivity, and understand various information in order to let robots adapt to and make smart decisions in complex environments.