A foldable small-scale soft electromagnetic robot for multimodal navigation in confined and unstructured environments

Abstract

Multimodal locomotion is crucial for an animal’s adaptability in unstructured wild environments. Analogously, in the human gastrointestinal tract, characterized by viscoelastic mucus, complex rugae, and narrow openings like the cardia, multimodal locomotion is also essential for a small-scale soft robot to conduct tasks. Here, we introduce a small-scale, compact, foldable, and robust soft electromagnetic robot with more than nine locomotion modes designed for such a scenario. Featuring a six-spoke elastomer body embedded with liquid metal channels and driven by Laplace forces under a static magnetic field, the robot is capable of rapid transitions ( < 0.35 s) among different locomotion modes. It achieves exceptional agility, including high-speed rolling (818 mm/s, 26 body lengths per second), omnidirectional crawling, jumping, and swimming. Notably, the robot can fold to reduce its occupied volume by 79%, enabling it to traverse confined spaces. We further validate its navigation capabilities in unstructured environments, including discrete obstacles, viscoelastic gelatin surfaces, viscous fluids, 3D-printed gastric surfaces, and a porcine gastric environment. This system offers a versatile strategy for developing high-mobility soft robots for future biomedical applications.

The authors present a small-scale, foldable soft electromagnetic robot that features nine locomotion modes and a high rolling speed, enabling navigation through confined and unstructured environments for potential biomedical applications.

Introduction

Multimodal locomotion enables organisms to adapt to complex environments, thereby enhancing their survival chances^1–7. Inspired by these biological systems, researchers have developed various small-scale robots capable of multimodal locomotion to achieve functional versatility^8–16 with applications in path navigation^17,18, inspection and detection^19,20, and targeted drug delivery^21–23. These capabilities are particularly critical for operation within confined and unstructured settings, such as the human gastrointestinal tract^24,25, which poses significant challenges: viscoelastic mucus, complex rugae, and narrow openings (e.g., the cardia) often constrain the mobility of existing soft robots. Current devices frequently suffer from limited adaptability to terrain irregularities, leading to functional failure during abrupt transitions or posture changes like flipping^9–11,15,17. Therefore, developing a soft robot that combines a configurable structure with high mobility is essential for navigating such unstructured environments²⁴. Achieving this goal requires a synergy of material selection and structural design.

To facilitate the miniaturization and structural simplification of small-scale robots, diverse smart materials offer direct energy conversion without requiring complex and bulky transmission mechanisms, including shape memory alloys^10,26,27/polymers²⁸, pneumatic systems^29,30, stimuli-responsive polymers^31–34, dielectric elastomers^35–37, and magnetic composites^8,14,23,38, have been extensively investigated. However, their widespread application in confined and unstructured environments, such as the human gastrointestinal tract, is often hindered by intrinsic limitations, ranging from slow response speeds and bulky tethering systems to safety concerns regarding high voltage or difficulties in localized control. Soft electromagnetic actuators (SEMAs) driven by Laplace forces offer a compelling alternative, featuring millisecond-scale response, programmable deformations, and precise control^39–41.

Beyond materials, structural design is equally decisive. Crawling robots, typically employing bipedal or multi-legged structures, demonstrate excellent maneuverability. For instance, a bipedal robot developed by Mao et al. achieved a remarkable speed of 70 BL/s¹¹, while a bio-inspired multi-legged design from Ji et al. exhibited omnidirectional mobility¹⁷. However, neither design is suitable for performing in complex environments. Swimming robots typically feature streamlined, elongated bodies for propulsion in aquatic environments⁴², yet lack terrestrial mobility. Jumping robots, which utilize bistable mechanisms⁴³ or elastic energy storage systems⁴⁴ for rapid energy release, are effective for obstacle traversal but difficult for sustained locomotion. Rolling robots offer energy efficiency on flat surfaces, serving as a pathway to enhance mobility for robots, but typically struggle in unstructured environments and lack integration with other modes^{30,32,34,36,37,45,46}. For instance, a humidity-responsive robot developed by Xu et al. achieved a slow rolling speed of 5.8 BL/s³². In contrast, a dielectric elastomer-based design by Li et al. reached only 0.98 BL/s on flat terrain³⁶. While recent electromagnetic rollers achieved high speeds of 14 BL/s⁴⁷, they remain limited to only a single locomotion mode, failing to address the complex navigational needs of the unstructured environment.

Drawing inspiration from the Golden Wheel Spider and the Moroccan Flic-flac Spider, which utilize postural adjustments for rapid multimodal transitions between different modes^48,49 (Supplementary Fig. 1a), we developed a small-scale (typically an order of magnitude of 10⁻² to 10⁻³ m), foldable, and multimodal soft electromagnetic robot (M-SEMR) designed for robust navigation in unstructured environments like the human gastrointestinal tract (Fig. 1a). The M-SEMR features a six-spoke elastomer structure embedded with liquid metal channels, enhancing its ability to traverse undulating and viscous surfaces. Driven by Laplace forces within a global static magnetic field, the robot is designed to fold significantly to traverse narrow openings like the cardia, and subsequently unfold in the stomach to deliver drugs to target areas. The M-SEMR can dynamically adjust among nine gait modes and postures, providing the richest motion repertoire among small-scale robots and surpassing existing soft robots in maneuverability (Fig. 1b, c), and showing a relatively low consumption evaluated by the cost of transport (value of 27.3, Supplementary Fig. 2). Specifically, the M-SEMR utilizes two distinct postures, standing and lying, to enable modes, including rolling, crawling, walking, jumping, swimming, steering, backing, and posture switching, all achieved through control strategies rather than extra reconfigurable actuators. Notably, the M-SEMR can achieve a maximum rolling speed of 818 mm/s, complete a posture switch in 0.35 s, and perform in-place turning maneuvers. Furthermore, the robot demonstrated amphibious capabilities, including underwater deployment and underwater locomotion. The M-SEMR is also driven across various unstructured environments, including viscoelastic gelatin surfaces, discrete stepped surfaces, viscous fluids, and a 3D-printed gastric surface, providing a versatile solution for traversing viscoelastic mucus and complex rugae.

Fig. 1. Multimodal locomotion and performance of the M-SEMR.

a The M-SEMR can be compactly folded into a cylindrical configuration to reduce its occupied volume. Upon release, it can switch actively between two postures. The standing posture supports rolling or walking in three modes. The lying posture supports U-actuation, V-actuation, or crawling with three dual-module coordinated actuation modes. b Comparison of locomotion modalities between the M-SEMR and other reported soft robots with multimodal locomotion. Details are provided in Supplementary Table 1. c Relationship between the relative speed and body length of the M-SEMR, insects, mammals, rolling soft robots, and multimodal soft robots. The tethered robots are marked with a dashed circle. Details are provided in Supplementary Table 2.

Results

Design and fabrication

The M-SEMR is a fully soft and modular robotic system actuated in a static magnetic field, adopting a spoke configuration composed of six identical driven modules, which are soft electromagnetic actuators³⁹, made of a molded elastomeric shell embedded with liquid metal channels as soft conductive coils (Fig. 1a). Each module can be driven individually by Laplace forces resulting from a sequence of currents applied to the soft conductive coil. The modules are assembled using mortise and tenon joints, which minimize redundant components and overall system mass (Supplementary Fig. 1b, c). The entire M-SEMR weighs only 3.84 g and exhibits an internal resistance of 0.19 Ω per module.

To ensure the structural robustness of M-SEMR during its rolling, we characterize the mechanical response of the elastomer (Supplementary Fig. 3a) and conduct a buckling analysis of a single module. Because the whole body of the rolling M-SEMR is periodically supported by a single module (Supplementary Fig. 3b), this may result in a buckle that leads to motion destabilization. Considering the buckling analysis (Supplementary Fig. 3c), flexibility, and fabrication of the module, the dimensions of the module are set to 16 mm × 14 mm × 2 mm (Supplementary Fig. 3d). The cross-section of the liquid metal channels is a square with dimensions of 0.5 mm × 0.4 mm. The gap between adjacent liquid metal channels is 1 mm to ensure electrical insulation and structural strength of a single module. The number of modules in an M-SEMR affects its rolling performance and environmental adaptability. Adding more modules to one M-SEMR improves the robot’s peripheral envelope roundness, making it closer to an ideal circle (Supplementary Fig. 4a), thereby enhancing the smoothness and continuity of rolling (Supplementary Fig. 4b). Thus, too many modules reduce the foldability and obstacle clearance of the rolling M-SEMR. Here, we balanced these competing criteria, which led to the selection of a six-module design (Supplementary Fig. 4c). For actuation, the M-SEMR operates within a uniform magnetic field (Supplementary Fig. 5) and is powered by a designed pulse width modulation (PWM) controller (Supplementary Fig. 6). Each module can be independently actuated using square wave current with adjustable frequency and duty cycle (Supplementary Fig. 7), enabling diverse deformation, including bending, twisting, compression, and extension (Supplementary Fig. 8). Experiments demonstrate that such a design of M-SEMR supports fast multimodal locomotion and adaptability to unstructured environments.

Rolling

By maintaining a constant current and magnetic field direction while actuating the modules of the rolling M-SEMR sequentially, the resulting Laplace force vector is dynamically modulated to achieve targeted actuation across specific modules. Based on the dynamic analysis and actuation strategy (Supplementary Figs. 9–11), fast rolling of the M-SEMR requires shortening its startup time, defined as the minimum pulse duration needed to transition from the initial state to state 2. A simple solution is to increase the current amplitude. As an example, when the pulse current increases from 1 A to 1.5 A, the startup time reduces from 40 ms to 20 ms (Fig. 2a), with a 30% decrease in the transition time between states 2 and 3 (Fig. 2b). The motion process of the rolling M-SEMR involves complex dynamic and finite deformations, making it challenging to describe with a simple theoretical model. Here, instead, we simplify the M-SEMR to a rigid body and build a theoretical model to analyze its behavior during startup time. Details about this theoretical model are provided in the Supplementary Materials. Based on the theoretical model, the startup time is 38 ms, higher than the 20 ms obtained from experiments (Supplementary Fig. 12). From experimental observation in Fig. 2a and Supplementary Fig. 13, the soft module C2 first bends at high speed, so that the Laplace force performs much more work in a short time than predicted by the theoretical model. To better understand this, we conducted both multibody dynamics analysis (Supplementary Fig. 14a and Supplementary Table 4) and finite element analysis (FEA, Supplementary Figs. 15, 16). In multibody dynamics analysis, the rigid-body M-SEMR yielded a startup time of 35 ms, agreeing with the result from the theoretical model. The semi-rigid M-SEMR yielded a startup time of 24.5ms. Results from the multibody dynamics analysis of the two types of M-SEMR demonstrated that the semi-rigid version generates over an order of magnitude more work than the rigid version within a 20 ms actuation (Supplementary Fig. 14b). The FEA further revealed the underlying mechanism that the work performed by Laplace forces is converted into both kinetic energy and strain energy within the modules (Supplementary Fig. 15). While the FEA predicted a faster 13-ms startup time due to several unaccounted real-world complexities, it still confirmed that the soft structure performs significantly more work than a rigid counterpart. These consistent findings from both multibody dynamics analysis and FEA highlight the importance of structural compliance in energy transfer and agility for M-SEMR.

Fig. 2. Performance of the rolling M-SEMR.

The amplitude of the current is 1.5 A. a Snapshots of the posture of M-SEMR subjected to a 20-ms pulse current. b Time costs for the M-SEMR to reach different states with various pulse durations. The robot does not reach state 2 with a 10 ms pulse. c Maximum rolling speed versus actuation frequency of the M-SEMR subjected to a 1.5 A square wave current. d Success rate of the M-SEMR rolling from initial state to state 3 as a function of actuation frequency. There are 40 measurements conducted for each actuation frequency distributed across 2 prototypes. e Snapshots of the rolling M-SEMR subjected to a sequence of 60 Hz square wave currents, followed by the global actuation signals. The peak speed is 818 mm/s (26 BL/s). f Trajectories (green lines) of the M-SEMR’s mass center during continuous rolling on various surfaces, along with corresponding max. speeds illustrated on the right side. All error bars represent the standard deviation of five measurements.

To achieve rapid and continuous rolling of the M-SEMR, square-wave currents were sequentially applied to individual modules from C2 to C1 in an anticlockwise order (Supplementary Movie 1). We defined the actuation signals into two types: global actuation signal and local actuation signal (Supplementary Fig. 6b). The specific frequencies used for each type in different experiments are detailed in Supplementary Table 5. Supplementary Fig. 17 shows representative displacement, velocity, and acceleration versus time curves of the M-SEMR subjected to various actuation frequencies of square wave currents. The rolling speed increases with actuation frequency and reaches a maximum of 818 mm/s (26 BL/s) at 60 Hz (Fig. 2c and Supplementary Movie 1), representing the fastest rolling velocity reported for small-scale soft robots. However, the pulse duration of the 60 Hz square-wave current (the global actuation signal) is only 16.7 ms. This duration is shorter than the M-SEMR’s startup time, making it unsuitable for the continuous rolling of the M-SEMR. As a result, the rolling success rate decreases sharply to 10% (Fig. 2d). To overcome this issue, we implemented a hybrid actuation strategy combining a short time of low-frequency actuation with a subsequent high-frequency actuation (Supplementary Fig. 18). This approach substantially improved the rolling success rate to 80% (Fig. 2d), preserving a high rolling speed of 698 mm/s (see the data of 60 P in Supplementary Fig. 17d). A sequence of snapshots of the rolling M-SEMR is shown in Fig. 2e. During such rapid rolling, the robot exhibits intermittent jumping, which improves its adaptability to environments with unstructured surfaces. Balancing the rolling speed and locomotion stability from Fig. 2d, a global actuation frequency of 40 Hz was selected for the following demonstration if there is no extra statement. With such an actuation frequency, the M-SEMR demonstrated robust rolling across a variety of substrates, including stainless steel, silicone, PMMA, polyurethane, sandpaper, and paper (Fig. 2f and Supplementary Movie 1). On all tested surfaces, the robot achieved high speeds exceeding 500 mm/s and a maximum of 643 mm/s on paper. The trajectory analysis, as shown in the displacement-time curves of Supplementary Fig. 19, demonstrates that M-SEMR achieves extra stable rolling on both polyurethane and PMMA surfaces with high consistency.

Walking and crawling

In addition to the rolling mode, the M-SEMR also exhibits three walking modes: walk-1, walk-2, and walk-3 (Fig. 3a–c and Supplementary Movie 2). In walk-1, only one top module, such as C1 or C2, of the M-SEMR is actuated. As an example, in Fig. 3a, C1 bends back and forth under the Laplace force that induces the rotation of the whole body, producing an inchworm-like gait to the left side with an average speed of 16 mm/s (Supplementary Fig. 20a). If only C2 is actuated, the M-SEMR will walk towards the opposite side so that the walking direction of the M-SEMR can be well controlled. In walk-2, only one bottom module is actuated, such as C5 in Fig. 3b. When C5 is driven by the Laplace force and bends up forward, C4 is the only support for the M-SEMR, which induces the bending of C4 and the forward movement of the whole M-SEMR. Once the Laplace force is removed, the M-SEMR tends to return to its standing posture (Supplementary Fig. 20b). With such locomotion, the M-SEMR can walk at a speed of 24.26 mm/s. Similar to walk-1, the M-SEMR can switch its walk direction by switching the actuated bottom module. In walk-3, the top two modules are actuated simultaneously using a pair of square wave currents, which have the same amplitude but a phase difference of half a period. The mechanism of this mode is similar to that of walk-1. In such a mode, the robot can walk at a speed of 23.76 mm/s (Fig. 3c and Supplementary Fig. 20c), which is higher than that of walk-1.

Fig. 3. Characterization of walking and crawling M-SEMR.

The driven current in this panel is 1.5 A. Displacement versus time curves of M-SEMR adopting three walking strategies: (a) walk-1, (b) walk-2, and (c) walk-3 in a standing posture with highlighted modules actuated (bottom right corner). The slope of the linear fit corresponds to the average speed. d Dynamic response of the module torsion angle to single pulse-width durations in the M-SEMR lying posture. Each data point is obtained from 20 measurements. Three dual-module coordinated actuation strategies: (e) O-actuation, (f) M-actuation, and (g) P-actuation in the M-SEMR lying posture. Trajectories during linear motion, displacement-time curves of linear movement, and turning trajectories for the M-SEMR under different driving strategies. Yellow-highlighted modules indicate the current actuation modules. Ro, Rm, and Rp represent the turning radii corresponding to their respective actuation strategies. h The M-SEMR completes the ZJU trajectory using the P-actuation strategy.

The rolling and walking modes are based on the bending deformation of modules. The M-SEMR can also use the twist of modules to achieve crawling locomotion. Different from the standing posture of previous locomotion modes, crawling requires a lying posture. The relationship between the maximum torsion angle and the duration of a pulse current was experimentally characterized. As shown in Fig. 3d, the torsion angle increases with the pulse current duration and reaches a maximum of 41° once the duration exceeds 10 ms. Accordingly, a 10 ms actuation pulse was chosen to ensure rapid and sufficient deformation for fast straight movement. A 4 ms pulse was chosen for a slight deformation, which is beneficial for M-SEMR to turn in place. As Fig. 1a shows, for U-actuation, when only a module is actuated, the M-SEMR can move a short distance. Selective activation of modules C1 through C6 enables omnidirectional crawling, and the robot retains mobility even with partial actuator failure, demonstrating strong fault-tolerant capabilities.

To demonstrate the maneuverability of the M-SEMR in its lying posture, we developed three dual-module actuation strategies, termed O-, M-, and P-actuation (inspired by the positional isomers of xylene, Fig. 3e–g), which depend on the positions of activated modules. The fundamental principle in all strategies is the polarization difference between the two activated modules. When the two modules are driven with an opposite polarization, a combined force propels the robot forward or backward in a linear path. By selectively activating different pairs of modules using these three strategies, the M-SEMR can achieve six-direction crawling. When the two modules are driven with an identical polarization, it creates a torque that causes the robot to rotate clockwise or anticlockwise (Supplementary Fig. 21). We implemented these two pairs of driving currents with three dual-module actuations. O-actuation (Fig. 3e): Two adjacent modules (such as C2 and C3, 60° apart) are activated. The angle between their driving forces is 120° (opposite polarization) or 60° (identical polarization), resulting in a slow linear forward crawling speed of 1.84 mm/s and a minimum turning radius of 6.4 mm (Supplementary Movie 3). M-actuation (Fig. 3f): Two modules separated by 120° (such as C2 and C4) are activated. This changes the angle between the forces to 60° (opposite polarization) or 120° (identical polarization), yielding a higher speed of 5.21 mm/s and a smaller turning radius of 3.7 mm (Supplementary Movie 3). P-actuation (Fig. 3g): Two opposing modules (180° apart, such as C2 and C5) are activated. The angle between the forces becomes 0° (opposite polarization) or 180° (identical polarization), enabling the highest speed of 10.51 mm/s and in-place rotation (Supplementary Movie 3). We demonstrated this maneuverability by using an open-loop strategy to trace a ZJU trajectory (Fig. 3h and Supplementary Movie 3), demonstrating forward/backward motions and clockwise/anticlockwise rotations of the M-SEMR.

Posture switching

In previous sections, we have introduced nine different types of locomotion modes of M-SEMR, which can be divided into two groups according to the corresponding postures, lying and standing. To freely switch among all nine locomotion modes, it is important to find a control strategy that enables the M-SEMR to change its posture easily (Fig. 4a). From experimental observation, the rolling M-SEMR enters an unstable oscillatory state under high-frequency actuation. When the current is cut off in such an unstable state, the M-SEMR naturally falls into the lying posture by gravity in 350 ms (Supplementary Fig. 22 and Movie 4). From experiments, we also observe that the M-SEMR can switch from the lying to the standing posture through a sequence of different driven currents. During this process of standing up, the torque from the Laplace force plays an important role, which overcomes the gravitational torque, elevating the center of mass (CoM) of the M-SEMR. A detailed analysis shows that each module (C1-C6) sequentially receives a 50-ms pulse. Initial activation of C1 establishes a ground contact at C5 and C6, forming a lever mechanism with the distal end serving as the pivot. Subsequent activation of C2 dynamically shifts the contact points to C6 and C1. After 300 ms, the robot re-establishes the standing posture (Supplementary Fig. 23). Compared with the transition from standing to lying, the transition from lying to standing involves more complex dynamics. Kinematic decoupling identifies two primary components:(1) sequential rotational actuation of individual modules and (2) elevation of the CoM through coordinated body rotation about distal contact points. The first component, induced by signal transitions, contributes minimally to net mechanical work and can be neglected. The second component is modeled as a module transitioning from vertical to horizontal, allowing magneto-mechanical analysis of torque balance and energy transfer (Supplementary Fig. 24a–d). Analytical results demonstrate that this rotational elevation reduces instantaneous power by elongating the movement path of the CoM while maintaining constant total work, thereby facilitating an energy-efficient transition (Supplementary Fig. 24e, f). In Fig. 4b, we plot the transition time from lying posture to standing posture as a function of actuation frequency, indicating that the transition time increases as the frequency exceeds 20 Hz. In addition, it is difficult for the M-SEMR to switch from lying to standing posture at frequencies above 60 Hz (such as 70 Hz).

Fig. 4. Posture switching and navigating through a confined space of the M-SEMR.

a Active posture switching of the M-SEMR in steps. b Relationship between transition time and actuation frequency during the M-SEMR’s posture switching from lying to standing posture. c An obstacle is placed above the M-SEMR’s forward path. d The M-SEMR switches to the lying posture. e The M-SEMR crawls through the passageway. f The M-SEMR crawls a designated distance away from the passageway. g The M-SEMR switches back to the standing posture. h The M-SEMR continues to roll. i Robustness test showing the resilience of M-SEMR upon external loading. The robot resumes rolling after being compressed and fully flattened. j Locomotion of the M-SEMR under compression in a narrow space.

To demonstrate the advantages of posture switching of the M-SEMR, we designed a task for the M-SEMR to navigate in a height-limited environment (Fig. 4c–h and Supplementary Movie 4) in 46 s, utilizing two locomotion modes and two posture-switching actions. A rectangular obstacle was set, leaving a passageway with a height of 26 mm, to obstruct the locomotion of the M-SEMR toward the right side. It needs to be pointed out that the height of the M-SEMR is 27 mm in the standing posture, and the maximum height of the rolling M-SEMR is 31.5 mm (Fig. 4c). Considering this, the M-SEMR cannot cross through the 26 mm passageway by rolling. Instead, the M-SEMR switched the posture from standing to lying in 0.5 s, actuated at 67 Hz, resulting in the height of the M-SEMR changing to 16 mm (lower than the passageway) (Fig. 4d). Once the transition was complete, the M-SEMR switched to the U-actuation mode and crawled through the passageway in 29 s, actuated at 16.7 Hz (Fig. 4e, f). Upon reaching the right side, the M-SEMR switched to its standing posture from a lying posture at a global actuation frequency of 20 Hz (Fig. 4g). Then, subjected to a 50 ms pulse, the M-SEMR started a first rolling process (Fig. 4h). A significant portion of the 46 s task time is due to the manual switching of control commands. This duration could be substantially reduced by implementing automated control. Through this task, the M-SEMR demonstrated its adaptability in a height-limited environment, thereby providing a robust approach for navigation in confined and unstructured environments.

The robustness of robots, encompassing both fault-tolerant and fail-safe strategies, is essential for their survival. The M-SEMR presented here exhibits remarkable robustness by integrating soft materials with a simple structure. As demonstrated in Fig. 3, the M-SEMR can maintain locomotion even if only a single module is functional, exhibiting exceptional fault tolerance. Experimentally, we also confirmed its body resilience to physical damage. The M-SEMR was compressed to 6 mm (20% of its height) by a load and fully recovered its functionality after the load was removed (Fig. 4i, Supplementary Fig. 25 and Supplementary Movie 4). To further demonstrate its capabilities, we tested its locomotion under sustained compression in a narrow space where standard movement was impossible (Fig. 4j and Supplementary Movie 4). Here, we designed two signals to selectively activate module pairs (C2 & C4; C1 & C5), inducing periodic bending that propels the robot forward (Supplementary Fig. 26).

Underwater locomotion

In addition to navigating on land, the M-SEMR can also realize underwater locomotion. Here, considering the future practical application of the M-SEMR in the gastrointestinal tract, we also demonstrated a deployment approach, as shown in Fig. 5a, where the M-SEMR (31.5 mm in diameter) is folded into a compact cylindrical configuration (14.5 mm in diameter) and packaged within a dissolvable polyvinyl alcohol (PVA) film. It is noted that the M-SEMR’s occupied volume is reduced to 21% in its folded state compared to its unfolded state. This approach enables the M-SEMR to cross a narrow opening, such as the gastric cardia, and then unfold autonomously in a water environment due to the dissolution of the PVA film. Figure 5b and Supplementary Movie 5 demonstrate the experimental deployment process of the folded M-SEMR into water through a tube (18 mm in inner diameter), including releasing, dissolving, and locomoting. The dissolving of the PVA film induced the release of the stored elastic energy in the M-SEMR, which unfolded automatically. Subsequently, when subjected to a 5 Hz square wave current, the unfolded M-SEMR performed a rolling locomotion toward the right side.

Fig. 5. Release, locomotion, and function of an M-SEMR in water.

a Schematic illustration of the M-SEMR’s folding, encapsulation, and autonomous release. The folded M-SEMR is encapsulated with a dissolvable film. b Snapshots of the folded M-SEMR at three stages: releasing, dissolving, and locomoting. c Locomotion with V-actuation, posture switching, and rolling of the M-SEMR.

The water environment brings two new major factors in the locomotion of the M-SEMR, including the additional buoyancy force and the high drag force of water, which induce other experimental phenomena. Experimentally, we found that the M-SEMR can actively switch between its two underwater locomotion modes: rolling and V-actuation (Figs. 1, 5c, and Supplementary Movie 5). In these two modes, all modules of the M-SEMR are driven in turns with a square wave current. The locomotion modes are highly frequency-dependent. For a low global actuation frequency, 5 Hz or 10 Hz, the M-SEMR can only locomote by rolling (Supplementary Fig. 27). The M-SEMR has difficulty rolling underwater at frequencies above 10 Hz. At high frequencies, the drag of the water restricts the movement of the M-SEMR’s modules. This fluid resistance prevents the module’s mechanical response from synchronizing with the fast-changing actuation signals. If the M-SEMR is in a lying posture, it can automatically switch to a standing posture and continue to roll. For a high global frequency, such as 40 Hz, the M-SEMR can only locomote with V-actuation. If the M-SEMR is in a standing posture, it can automatically switch to a lying posture and continue to locomote with V-actuation. We also tested the M-SEMR with higher global actuation frequencies, from 50 Hz to 100 Hz with a step size of 10 Hz, which shows the same phenomenon (Supplementary Fig. 28 and Movie 5). We further quantified the performance of these modes. The M-SEMR’s posture transition time in water is reduced to 200 ms, 150 ms faster than that on land (Supplementary Fig. 27a). From our experiments, the M-SEMR achieves a maximum average rolling speed of 51.7 mm/s at a global frequency of 5 Hz (Supplementary Fig. 27c, d) and an average locomotion speed of 35.1 mm/s with V-actuation at 50 Hz (Supplementary Fig. 28).

To demonstrate the potential applications of the M-SEMR, we conducted two experiments as shown in Supplementary Fig. 29. Here, we designed a controllable drug release module (Supplementary Fig. 29a) that was integrated with the M-SEMR. This module consists of a chamber, platinum electrodes, and a cover with a microchannel in its center. The diameter of the microchannel is set to be 0.5 mm, which is small enough to seal the liquid drug in the chamber. In actuation, the electrodes are connected to a DC power supply for the electrolysis of the water in the chamber. Along with the electrolysis, oxygen and hydrogen accumulate around the electrodes, pushing the liquid drug out of the chamber through the microchannel in the cover. As shown in Supplementary Fig. 29b and Movie 6, upon crawling to the target area, the M-SEMR releases its liquid drug by activating the drug module with a 10 V voltage for 3 seconds. Following the release, the M-SEMR is controlled to mix the liquid drug with water. From the mechanism of the drug-release module, the drug delivery process can be well controlled for multiple releases. As evidenced in Supplementary Fig. 29c and Movie 6, the M-SEMR crawled to three positions and released drugs with various dosages, demonstrating its controllable release capability.

Locomotion in unstructured environments

Most soft robots are operated in confined and unstructured environments, where environmental variations significantly impact their locomotion performance. The M-SEMR, however, leveraging its actuation strategy and structural design, demonstrates environmental adaptability across diverse environments (Supplementary Table 5).

To evaluate the M-SEMR’s environmental adaptability, we conducted a series of locomotion experiments on various terrains (Supplementary Movie 7). Results show that the M-SEMR can cross a series of small wedge blocks with a base length of 20 mm and varying inclinations from 6° to 38°, in less than 0.6 s (Supplementary Fig. 30). In addition, the M-SEMR can cross a series of small rectangular obstacles with heights ranging from 1 mm to 10 mm. This maximum height corresponds to approximately 0.37 times the M-SEMR’s body height (Supplementary Fig. 31a). Besides, we built two periodic surfaces, a square shape and a sinusoidal shape, on which the M-SEMR can locomote easily, as demonstrated in Fig. 6a, and Supplementary Figs. 31b, 32. In addition, we prepared a 3D-printed stomach with a surface topography resembling the complex gastric rugae, where the M-SEMR easily traversed in 34.5 s in its lying posture and only 0.5 s in its standing posture, achieving a speed comparable to that on flat surfaces (Supplementary Fig. 33).

Fig. 6. Locomotion of M-SEMR in unstructured environments.

a Snapshot of the M-SEMR locomoting on a periodic square-shaped surface. b Snapshot of the M-SEMR locomoting on a viscoelastic gelatin surface. c Snapshot of the M-SEMR locomoting in a viscous fluid. d Snapshot of the M-SEMR locomoting from water to land. e Snapshot of the M-SEMR functioning in a porcine stomach (ex vivo). The scale bar is 10 mm.

Further, we created a soft and adhesive surface made of gelatin, on which the M-SEMR achieved an average speed of 226.2 mm/s at 20 Hz. This high speed is attributed to the M-SEMR’s spoke structure, which minimizes the contact area during rolling and thereby effectively prevents adhesion to the viscoelastic gelatin surface (Fig. 6b). And from our experience, most small-scale robots cannot even walk on such a viscous surface¹¹. Then, we tested the locomotion of the M-SEMR in a viscous non-Newtonian fluid made of yogurt solution (Supplementary Fig. 34). The viscosity of the yogurt solution is 800 mPa·s at 1 rad/s, which is higher than that of gastric mucus, about 85 mPa·s⁵⁰. In this viscous fluid, the M-SEMR, driven by a series of currents with a global frequency of 2 Hz, achieved a rolling speed of 34.5 mm/s. We also observed that the rolling locomotion was hindered at higher frequencies (such as 10 Hz) because of the high startup time of the rolling M-SEMR in viscous fluid (Fig. 6c). Apart from a solid surface or a liquid environment, the M-SEMR can roll from an aquatic to a terrestrial environment within 2 s, demonstrating its environmental adaptability (Fig. 6d).

Based on these results, we further demonstrated the M-SEMR’s performance in diverse complex environments with varying structures and materials to comprehensively evaluate its potential in applications (Supplementary Movie 8). For instance, we tested the M-SEMR within a real porcine stomach (ex vivo) filled with viscous fluid to mimic the gastric environment. In this setting, the initially folded M-SEMR successfully unfolded after the encapsulating film dissolved. It navigated across complex gastric rugae to reach a target area. Finally, the M-SEMR released the liquid drug in a controlled manner (Fig. 6e).

Discussion

In this work, we present a small-scale, compact, foldable, and robust soft electromagnetic robot with nine locomotion modes, which owns the most abundant modes among small-scale robots, actuated exclusively by Laplace forces within a static magnetic field generated by an electromagnet (such as an MRI machine, which can greatly improve the mechanical efficiency of the M-SEMR³⁹) and constructed using a modular multi-module design. The developed strategy in posture switching, between lying and standing in less than 0.5 s, enables the M-SEMR to change its locomotion modes among rolling, three walks, three crawls, and locomotion with U, V-actuation freely. In the standing posture, the M-SEMR achieves a top rolling speed of 818 mm/s (26 BL/s), representing the fastest reported rolling performance among small-scale soft robots. It also maintains a speed of up to 500 mm/s across various substrate surfaces, including stainless steel, silicone, PMMA, polyurethane, sandpaper, and paper.

Notably, the robot incorporates a foldable design that reduces its occupied volume to 21% of its original size, allowing it to access narrow openings and subsequently deploy for task execution. The M-SEMR also shows high robustness in the experiments, enabling it to maintain locomotion with only a few modules and exhibiting resilience to physical damage, such as a high degree of compression. Experiments show that the M-SEMR can cross various unstructured environments, including discrete obstacles, viscoelastic gelatin surfaces, viscous fluids, and 3D-printed gastric surfaces. This performance across diverse environments stems from the robot’s ability to switch locomotion modes and postures adaptively. Moreover, we conducted an experiment where the robot locomoted in a porcine stomach and released drugs at a targeted position, highlighting its feasibility in a gastric environment.

Despite these advancements, the current system faces several technical challenges. First, the molding method and handmade fabrication limit further miniaturization, which can be solved with advanced technology such as high-precision multi-material 3D printing. Second, the current M-SEMR is still a tethered version, and a custom design of micro-power and controller still needs to be developed. Third, even though we have demonstrated a series of locomotion modes and control strategies for the M-SEMR, we believe there is still room to explore. Fourth, although this study demonstrates enhanced environmental adaptability, the dynamical response of the M-SEMR in actuation requires further systematic investigation for better performance.

Crucially, the M-SEMR achieves multifunctionality without compromising the performance of individual functions, allowing a single robot to be repurposed for distinct tasks without the need for replacement²⁴. The six-spoke morphology and folding capability of the M-SEMR provide generalizable mechanical strategies for the miniaturization and volume reduction of other multimodal robots, demonstrating a promising solution for future in vivo applications, including drug delivery, minimally invasive surgery, and health monitoring.

Methods

Materials

The M-SEMR’s architecture integrates six modular modules, each consisting of geometrically patterned soft elastomeric shells encapsulating liquid metal. The liquid metal, Galinstan (Dongguan Huatai Metal Material Technology Co., Ltd.), consists of gallium, indium, and tin with a mass ratio of 68.5:21.5:10. The elastomers were prepared by mixing Ecoflex-30 (Smooth-On Inc) and PDMS (Sylgard 184, Dow Corning) solution at a mass ratio of 10:1. A silicone adhesive (Sil-Poxy, Smooth-On) was used to bond the elastomers. The central support column in M-SEMR was fabricated from PLA (PLA, Bambu Lab). In addition, a drug-release module was integrated into the robot system. This module was fabricated using 3D printing with a photopolymerizable resin (Rigid 10k resin, Formlabs) and features embedded inert platinum electrodes for water electrolysis.

Fabrication

Step 1, mixing: The Ecoflex-30 solution was prepared by mixing Part A and Part B with a mass ratio of 1:1. Separately, the PDMS solution was formulated with a 10:1 mass ratio of base monomer to curing agent. These two solutions were then thoroughly mixed at a mass ratio of 10:1 (Ecoflex-30: PDMS) using a planetary centrifugal mixer (BHZ-300, Beihong) at 2000 rpm for 3 minutes. Step 2, molding: The resulting mixture was poured into a PETG mold fabricated via a 3D printer (X1 Carbon, Bambu Lab). The mold filled with solution was placed in a vacuum chamber (100 mbar) for degassing for 3 minutes. An acrylic plate, which is fixed with clips, was used to remove excess material and cover the mold. Step 3, curing: The mold was placed in an oven at 50 °C and cured for 25 min to form the elastomeric substrate. Step 4, demolding: The cured elastomers were carefully separated from the mold using tweezers, yielding two types of components: elastomer-1 (with predefined liquid metal channels; cross-section 0.5 mm × 0.4 mm) and elastomer-2 (with a flat surface). Step 5, forming sealed modules: A silicone adhesive (Sil-Poxy, Smooth-On) was applied to the flat side of elastomer-2, which was then aligned and bonded to elastomer-1. The assembly was subsequently placed in an oven at 60 °C for 3 min to complete the bonding process, resulting in individual and sealed elastomeric modules with empty channels. Step 6, liquid metal injection and wiring: For each module, liquid metal (Galinstan) was injected into its channels using a syringe with a 30 G needle. Electrical wires (diameter: 0.1 mm) were then connected to the injected liquid metal to form a functional actuation module (Supplementary Fig. 1b). To ensure robust attachment between the rigid wires and the soft elastomeric modules, the connection points were sealed with a silicone adhesive. The procedure was repeated for all six modules (Supplementary Fig. 1c). Step 7, final assembly: The six functional modules were assembled into a multi-layer structure. The silicone adhesive was applied to the bonding interface of one module, which was then aligned and attached to its adjacent module. This bonding process was repeated, and the entire assembly was cured at 60 °C to form the final six-layer elastomeric structure. Finally, the central support column, a 3D-printed PLA hexagonal prismatic support, was embedded at the center of the structure during the final assembly steps, completing the fabrication of the M-SEMR. More details can be found in the Supplementary Materials (Supplementary Fig. 1).

Control and power systems

The control system includes a microcontroller (Uno Rev3, Arduino) and a 16-channel PWM controller (PCA9685, Adafruit), which provides twelve independently programmable PWM signals for this setup. The drive system consists of six H-bridge motor driver modules (L298N, ST), each powered by a dedicated, independently regulated DC power supply (IT6833A, ITECH; or UDP6720, UNI-T). Every H-bridge module, controlled by two PWM signals, can output a programmable bidirectional current with a magnitude of up to 2 A.

Numerical simulation

The results of multibody dynamics analysis were obtained using the commercial software Adams. To capture the structural compliance, the main body is modeled as six interconnected rigid modules coupled via an equivalent spring-damping system. The flexibility of semi-rigid M-SEMR is obtained by fitting the stiffness coefficient and a damping coefficient. A constant Laplace force is applied to the liquid metal channel of module C2 to drive the rolling motion. The finite element analysis was performed using the commercial software Abaqus (C3D8R elements, explicit solver). We describe the mechanical behavior of both the elastomer and the internal liquid metal channel using a nearly incompressible linear elastic model. The elastomer’s Young’s modulus was calibrated based on experimental bending data. As it is not possible to apply Laplace forces directly in Abaqus, we implement the force as a surface traction on the liquid metal channels via a user-defined subroutine. The dynamic loading condition is executed sequentially: applying gravity to establish contact, introducing programmed surface traction for actuation, and removing the traction to allow structural recovery. Details are provided in the Supplementary Materials.

Trajectory tracking and speed measurement

The locomotion performances of the M-SEMR were recorded using different cameras for various experiments. (1) For rolling performance tests, a high-speed camera (ACS-1, NAC Image Technology) was used, recording at 2000 fps. The central black marker on the M-SEMR was tracked frame-by-frame using the software MVIEW (NAC Image Technology). A custom MATLAB script was used to calculate the instantaneous velocity considering the displacement across 10 frames (corresponding to a 10 ms interval) around the time point. (2) Walking and crawling motions were recorded using a high-speed camera (ACS-1) and a miniature camera module (DH-4k-003, Shenzhen Donghai Home Technology Co., Ltd.), respectively, both at 59.9 fps. The M-SEMR’s position was extracted using the software Tracker (Version 6.1.2, Open Source Physics). The velocity was calculated as the average speed over a 33-ms interval around each recorded time point. (3) Underwater locomotion was recorded at 59.9 fps using a mirrorless camera (Z5, Nikon). Velocity was processed using the same method as depicted in (2). (4) Off-road locomotion in complex environments was recorded at 59.9 fps. Two cameras, the high-speed camera (ACS-1) and the mirrorless camera (Z5, Nikon), were used simultaneously to capture from different perspectives. Velocity analysis was performed following the method described in (2). In Fig. 3, and Supplementary Figs. 17, 19, 27, 28, and 35, we plotted displacement and velocity as functions of time of M-SEMR’s locomotion in diverse environments.

Author contributions

G.M. and Z.L. conceived the project. Z.L. and X.Z. fabricated and characterized the M-SEMR. Z.L. developed the theoretical and numerical model. X.L. and X.Z. built dynamic simulations for the robot model. Z.L., G.M., and X.Z. designed the experiments. Z.L. and G.M. designed the figures and movies. M.Z. and X.S. assisted with the experiments. Z.L., X.Z., and G.M. analyzed the results and prepared and revised the manuscript. M.Z., X.S., X.L., Y.L., and S.Q. reviewed and edited the paper. G.M. and S.Q. supervised the research.

Peer review

Peer review information

Nature Communications thanks Dimuthu D.K. Arachchige, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the National Key R&D Program of China (grant no. 2023YFB4704700 to S.Q.), the National Natural Science Foundation of China (grant nos. 12372167 to G.M. and 12321002 to S.Q.), the Zhejiang Provincial Natural Science Foundation of China (grant no. LD26A020001 to G.M.), and the 111 Project of China (grant no. B21034 to S.Q.).

Data availability

All data needed to evaluate the conclusions of the work are available in the paper or the Supplementary Materials. Additional data related to this paper may be requested from the authors. Source data are available at https://doi.org/10.5281/zenodo.22023593.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-78006-w.