Asymmetries-induced nonlinear dynamic behaviors enable versatile modulation strategy for insect-scale robotics Authors: Yanting Gong, Zhongbin Guo, Kaijun Zhang, Dazhe Zhao, Tingxu Lan, Chengyue Lu, Yuan Ma, Junwen Zhong Published: 2025 ABSTRACT Developing versatile insect-scale robots that can handle various unstructured scenarios with a simple design remains a critical challenge. Here, we introduce a single electromagnetic actuator-driven insect-scale robot, developed by integrating geometric and mass asymmetries into a deformable hexagonal frame. This asymmetric configuration induces nonlinear dynamic behaviors in the robot, allowing it to perform a variety of complex actions (e.g., forward and backward locomotion, active flipping, and lateral movement) and respond to challenging scenarios (e.g., granular media and safe landing) through posture transitions. To facilitate the behavior modulation and design optimization of the proposed insect-scale robot, a nonlinear dynamic model is established to map the relationship between motion behaviors and input parameters. Last, an untethered prototype that can work in outdoor environments is constructed to further demonstrate the practical applications. INTRODUCTION Insect-scale robotics, capable of navigating through various unconstructed spaces, exhibit potential for performing tasks in challenging environments, such as narrow space exploration and urban disaster rescue (1–6). Developing versatile insect-scale robotics that can handle various unpredictable scenarios remains a challenge. Typically, a simple structural design results in lower manufacturing costs, higher reliability, and easier maintenance. However, this simplicity often restricts the versatility and adaptability of insect-scale robots, particularly when performing complex tasks in dynamic environments. Therefore, balancing the trade-off between design simplicity and functional diversity is essential for their practical applicability. Algorithm-based control approaches, such as the use of high–degree-of-freedom (high-DOF) platforms, are commonly used to achieve rich behavioral repertoires (7–9). These strategies typically require multiple actuators, complex coupling, and intensive real-time computation, which increase the overall hardware weight, software computational complexity, and calibration workload, making it difficult to miniaturize and lightweight the robot. For example, quadruped robots with more than 12 joints rely on detailed kinematic and dynamic modeling, along with reinforcement learning and trajectory optimization, to ensure gait stability (10). Alternatively, hardware-centric methods aim to enhance performance through structural innovation such as precise structural design (11–15) or modular assemblies (16–20). In structural design, a common approach is to integrate multiple actuation components into a single device—at least two. The differential modulation of these actuation components can not only readily alter the movement direction of microrobots but also enhance their adaptability in challenging environments (21–23). For instance, an insect-scale robot powered by three stacked dielectric elastomer actuators is developed, demonstrating autonomous navigation capability along printed paths (24). Similarly, Lai et al. (25) propose an architected design of passively morphing wheels integrated with electromagnetic motors. By the coordination of two distinct motors, the obtained robot achieves rapid movement on unstructured terrains. Despite these advancements, a key limitation of this strategy lies in the complexity of integrating multiple actuation components, which often leads to increased computational requirements, control complexity, and larger robot size. On the other hand, modular insect-scale robots offer an alternative approach, where units can be rapidly assembled and reconfigured to meet evolving demands. Specifically, Yoder et al. (26) introduce a class of electrohydraulic modules connected via magnetic interfaces. Depending on these reconfigurable actuation modules, they develop a robot capable of high-jumping and fast pipe crawling. Nevertheless, these designs often display limited maneuverability due to increased overall complexity induced by modular assembly. In addition, each module requires high degrees of consistency, imposing stringent demands on fabrication processes and time. Therefore, the inherent tension between simplifying the design of insect-scale robots and preserving their versatility remains a critical bottleneck. To address the above challenge, we propose an insect-scale robot that exhibits asymmetry-induced nonlinear dynamic behaviors, driven by a single electromagnetic actuator. The nonlinear dynamic behaviors are generated from integrating geometric asymmetry and asymmetric mass distribution into a deformable hexagonal structure. A general nonlinear dynamic model is established to describe the diverse behaviors and the fundamental modulation principles. Key advancements of this work include (i) the development of tethered and untethered insect-scale robots capable of diverse locomotion relying on asymmetries-induced nonlinear dynamic behaviors; (ii) the demonstration of superior terrain traversability across various terrains with a tethered prototype, including narrow pipes obstructed with small stones, high altitude with bluffs, and granular media; and (iii) the achievement of ultralow manufacturing cost and rapid assembly, with an untethered prototype costing ~$5.3 and requiring around 10 min for full assembly of both the robot body and circuit board. RESULTS Design and modulation principles As shown in Fig. 1A, the robot features a straightforward structure comprising a single electromagnetic actuator and a hexagonal support frame. The actuator is composed of a magnet and a planar coil, where the planar coil is fixed on the front end of the support frame and the magnet is placed at the center end. The two distal tips of the hexagonal frame serve as the robot’s “legs,” which contact the ground. When an alternating current passes through the coil, an alternating electromagnetic field is generated, resulting in a reciprocating vibration between the magnet and the coil. The primary design principle of the robot is to achieve a simple structural configuration that has abundant resonant modes and a low first-order resonant frequency, simultaneously. The multiple resonant modes enable robots to obtain diverse motion behaviors, while the lower first-order frequency facilitates large-amplitude responses with minimal excitation. The detailed parameters of the robot are presented in fig. S1 and tables S1 (tethered prototype) and S2 (untethered prototype). It is important to emphasize that the resulting robot displays asymmetry characteristics, including geometric asymmetry and asymmetric mass distribution. From the geometric asymmetry point of view, the front and rear legs of the robot are not parallel, displaying a tilt angle (θ) of 0.86° (Fig. 1B). The design of geometric asymmetry aims to achieve left or right turns in the proposed insect-scale robot. Mass asymmetry is introduced by placing more weight at the front end than the rear (Fig. 1C). The design of mass distribution asymmetry results in forward and backward movement. These asymmetries give rise to nonlinear dynamic behaviors of the robot, which can be modulated by tuning parameters, such as driving force, excitation frequency, friction coefficient, and the position of the center of mass (COM). For example, lower excitation frequencies result in negative displacements, while higher excitation frequencies produce positive displacements (Fig. 1D). Unless otherwise specified, the positive displacement (along the +y axis) is defined as forward movement. As a result, the obtained robot demonstrates enhanced maneuverability and terrain adaptability simultaneously. A robot can not only travel through challenging terrains more efficiently but also complete diverse movements in just one step. A conceptual application scenario is illustrated in Fig. 1E, where the robot operates in urban disaster settings that often include granular terrain, dead ends, and narrow spaces where turning is not feasible. Structural design and working mechanism of the insect-scale robot. (A) Schematic illustration of the tethered insect-scale robot. (B) Schematic illustration of the robot from a top view showing the geometric asymmetry design. (C) Schematic illustration of the robot from the side view showing asymmetric mass distribution. (D) Schematic illustration and modulation approach of the nonlinear dynamic system. (E) Conceptual application scenario for the robot. Optimization of key design parameters A most simplified insect-scale robot, consisting of two spring-damper components and a single electromagnetic actuator, is first developed to exhibit the potential of asymmetric configurations to help robots obtain diverse locomotion behaviors. As shown in Fig. 2A, the robot only has front and rear legs in contact with the ground. The planar coil is mounted at the front end, and the magnet is positioned centrally, resulting in a front-shifted COM of the total robot. Under this configuration, the robot moving inside a tube exhibits forward locomotion at a higher excitation frequency (~210 Hz), whereas a lower excitation frequency (~140 Hz) induces backward movement (Fig. 2B and movie S1). When an additional mass (about 85 mg) is attached to the rear end to shift the COM rearward, the motion direction is reversed: Forward and backward movements occur at lower and higher frequencies, respectively (Fig. 2C and movie S2). In general, at high excitation frequencies, the robot tends to move in the direction where the COM is located. To evaluate the influence of friction coefficients, the robot is tested under a constant voltage (1.8 V) and excitation frequency (180 Hz) on substrates with different friction coefficients. When the friction coefficient is ~0.46 (Sp-p10,000), forward motion is exhibited; when the friction coefficient is reduced to ~0.32 (glass), backward motion is shown (Fig. 2D). The above experiment results demonstrate that the most simplified robot with asymmetric configurations can obtain directional locomotion by adjusting the excitation frequency, the COM position, and friction. Optimization of key design parameters. (A) Schematic diagram of the simplest insect-scale robot. Optical photos showing the moving direction of the most simplified prototype moving inside a tube can be adjusted by the (B) excitation frequency, (C) COM position, and (D) friction coefficient. (E) Finite element method (FEM) simulation results showing the deformation of different hollow support frames under the same pressure, indicating that the hollow hexagonal column is easy to deform. (F) Changes in the displacement of different support structures under 1 N stress. The value is from the FEM. (G) Schematic diagram of the tilt angle, L1, and L2. (H) Relationship between excitation frequency and tilt angle. (I) Relationship between the vibrating amplitude of the hexagonal frame and L1/L2 ratio. PET, polyethylene terephthalate. To improve the movement performance of the robot, we optimize the design parameters of its critical component, i.e., support frame. We analyze the influence of geometric design on the mechanical response of the robots. Because reducing the structural stiffness of the support frame allows it to more effectively transmit the driving force generated by the electromagnetic actuator, thereby increasing the efficiency of converting input energy into mechanical motion. Figure 2 (E and F) shows the simulated deformation under a 1 N vertical load for different hollow support frames, including triangular, quadrangular, pentagonal, and hexagonal prisms. Compared to other geometries, the hexagonal frame exhibits larger deformation under vertical loading. This reduction in structural stiffness contributes to a lower first-order resonant frequency f=12πkm and improves electromechanical transmission efficiency. Thus, the hollow hexagonal prism is selected as the support frame of the robots. To explore the effect of geometric asymmetry, we investigate two key parameters: the tilt angle and the L1/L2 ratio. As shown in Fig. 2G, when the tilt angle alters from 0° to 3.43°, the first-order resonant frequency initially decreases and then increases. A minimum value is observed at 0.86° (Fig. 2H). We also examine the effect of the L1/L2 ratio on the vibration amplitude and motion stability of the robot. Simply, when the L1/L2 value is large, the robot’s vibration amplitude is large. However, excessive asymmetry (such as L1/L2 = 1) results in unstable motion. Besides, a lower ratio, such as L1/L2 = 0.3, produces insufficient vibration for effective motion. Based on this trade-off, we choose L1/L2 = 0.5 as the optimal value to balance locomotion efficiency and vibration amplitude (Fig. 2I). Further details can be observed in fig. S2. In this asymmetric hexagonal frame configuration, we further investigate the relationship between COM position and the robot motion direction. As shown in fig. S3A, the robot’s COM is changed by adjusting the magnet position. When the magnet is positioned above or below the center (center-top or center-bottom points), the robot only moves forward along the y axis over a wide range of excitation frequencies (100 to 900 Hz) (fig. S3, B and C). When the magnet is positioned to the left or right of the center (center-left or center-right points), the robot only exhibits rotational motion (fig. S3, D and E). Considering the multidimensional movement capabilities, the robot’s magnet is placed in the center. Furthermore, the electromagnetic actuator optimization of the robot is illustrated in note S1 and fig. S4. The trade-off between electromagnetic force and weight/size of the magnet and coil is considered. After optimization, the achieved tethered insect-scale robot (size: 12 mm by 12 mm by 18 mm, weight: ~810 mg) with asymmetric hexagonal frame and front-shifted COM position exhibits superior maneuverability, capable of performing diverse movements on the ground, such as forward movement, backward movement, forward-right turn, and forward-left turn (Fig. 3A and movie S3). We must emphasize that the actual motion mode of the robot is quite complex because the hexagonal structure is easy to deform. A complete forward or backward movement does not exist. When the turning radius is more than 50 BL, the behavior of robots is defined as straight movement. As shown in Fig. 3B, the straight movements are found at ~90 Hz (turning radius of ~200 BL) and ~270 to 330 Hz (turning radius of ~70 BL). In addition, the high excitation frequencies (>350 Hz) result in turning motions. When the excitation frequency is changed from 350 to 600 Hz, the turning radius of the prototype robot first increases and then decreases (fig. S5, A and B). The best angular velocity of ~300° s−1 at 400 Hz is achieved (movie S4). Figure 3C depicts the amplitude of the front and rear legs over time. The maximum amplitude of the front leg (~0.8 mm) is obtained at the first-order resonant frequency (~250 Hz). When the excitation frequency is >350 Hz, the displacement decreases to less than 0.2 mm. In general, if the moving amplitude of the robot decreases, then its moving speed will also decrease. Thus, the movement velocity of the robots should be the fastest at the first-order resonant frequency. However, at this operating frequency, antagonistic interactions between forward and backward movements hinder effective displacement of the robot. In addition, the increased vibration amplitude raises the likelihood of flipping. Thus, the maximum forward speed of ~19.5 BL s−1 is achieved at 330 Hz (fig. S5E and movie S5). Similarly, the fastest backward speed of ~9.5 BL s−1 is achieved at 290 Hz (fig. S5F and movie S5). The optimized robot displays a reversible moving displacement under different input voltages, as shown in Fig. 3D. For a fixed operating frequency, the robot moves backward at 0.7 V. When the input voltage is increased to 2.1 V, it moves forward (movie S6). The relationship between motion behaviors and input voltages is shown in Fig. 3E. When the input voltage exceeds 1.8 V, the initial backward motion changes to forward. The relationship between COM positions and the motion behaviors of robots is also explored, as illustrated in Fig. 3F. Initially, the robot moves toward one direction at 330 Hz. When an external mass of 4 g is placed at the rear end of the robot, the robot moves to the opposite direction (movie S7). Notably, the driving force and friction coefficient are unchanged. In this situation, the only change is the position of the COM, shifting from the front end to the rear end. In general, the basic working mechanism and motion behavior of the optimized robot prototype is consistent with that of the simplest robot prototype. Enhanced maneuverability in tethered insect-scale robot. (A) Optical photo showing a robot with hexagonal frame capable of backward movement, forward movement, forward-right turn, and forward-left turn. The adjustment strategy is to change the value of the excitation frequency. (B) Measurements of the movement speeds of a robot prototype (12 mm by 18 mm by 12 mm) and turning radius versus the excitation frequency (input voltage: 2.6 V). (C) Amplitude of front leg and rear leg of a robot as a function of the excitation frequency. (D) Optical photo showing the relationship between the movement of robot and the input voltage. (E) Moving speed as a function of input voltage. (F) Optical photos showing the influence of the COM position, (top) the COM position is positioned in the front end, and (bottom) the COM position is in the rear end. A series of high-speed images showing one bouncing cycle during the (G) forward movement and (H) backward movement of the robot. FEM simulation showing (I) the first-order and (J) fourth-order vibration modes of the robot. FEM simulation showing the displacements of different points in the (K) first-order vibration mode and (L) fourth-order vibration mode. All error bars represent the SD of three measurements. Moreover, to deeply understand the movement gaits, the real-time postures and positions of the insect-scale robot with hexagonal frame are recorded by a high-speed camera [sampling rate of 5000 frames per second (fps)] (movie S8). Taking forward movement as an example, at the initial stage (t = 0 ms), the rear and front legs stick to the substrate (Fig. 3G). Subsequently, the COM of the robot moves periodically along the y axis when an alternating current is applied. When the time reaches 2 ms, the rear leg sticks to the substrate, while the front leg is released. In this stage, the position of COM is increased along the z axis. After 6 ms, the rear leg leaves the ground with a front leg extension posture. The flying time remains about 1 ms. Because the position of COM is placed at the front end, the front leg always first touches the ground. The moving gait of the backward is similar to the forward (Fig. 3H). Contrary to it, the robot moves backward when the rear leg jumps above the ground and the front leg sticks. To explore the relationship between vibration modes and dynamic behaviors, finite element method (FEM) simulation is used. The detailed simulation and relative parameters used are shown in note S2 and table S3. The calculated first six orders of modes are illustrated in table S4. The first-order resonant frequency is ~259 Hz, and the maximum deformable displacement is ~0.8 mm at the front leg (Fig. 3I). In comparison, the fourth-order resonant frequency is ~529 Hz, exhibiting an obvious deformation at the rear leg (Fig. 3J). Moreover, the displacements of the four ground contact points of robots are recorded. Similar to the vibration mode, the displacement at the front end is larger than that at the rear end in the first-order vibration mode (Fig. 3K). In contrast, the fourth-order vibration mode exhibits a more obvious deformation at the rear end (Fig. 3L). Nonlinear dynamic modeling To explain the forward and backward movement of the designed robot and explore the relationship between input parameters and nonlinear dynamic behaviors, a simplified asymmetric mass-spring-damping model is established (Fig. 4A, note S3, and table S5). Lauga and Bartolo (27) demonstrate that by introducing a phase difference in the deformation period of a multiparticle system, time-reversal symmetry can be broken, thereby achieving cooperative motion of the particles. Although the principles governing fluid mechanics and ground locomotion of our insect-scale robot rather differ, the fundamental condition for producing a net propulsion is similar: introducing asymmetric configuration to break time-reversal symmetry. Considering this, the robot is simplified into three blocks with different masses, including the front end, center end, and rear end. The different ends are connected via two spring dampers (Fig. 4A). The dynamic behavior of the different ends is described by ordinary differential equations (ODEs){mfrontafront=−F−k(xfront−xcenter)−β(vfront−vcenter)−ffrontmcenteracenter=F+k(xfront−xcenter)+β(vfront−vcenter)−k(xcenter−xrear)−(vcenter−vrear)mreararear=k(xcenter−xrear)+β(vrear−vrear)−frear(1)where vfront, vcenter, and vrear are the relative moving speeds of the front end, center end, and rear end; β and k are the damping coefficient and spring coefficient; xfront, xcenter, and xrear are the positions of front end, center end, and rear end; F is the driving force; and ffront and frear represent friction force of front and rear ends (note S4). An iterative method, named the “fourth-order Runge-Kutta method,” is used to solve the numerical solution of ODEs. Moreover, the damping coefficient of the spring dampers is achieved by the logarithmic decay method (note S5 and fig. S6). The natural frequency of the system is obtained by solving the mass matrix and the stiffness matrix (note S6). The vibration mode (fig. S7) at natural frequency is evaluated by FEM simulations. The behaviors in the systems are determined by the displacement of the COM. Nonlinear dynamic modeling. (A) Schematic illustration of the model construction. (B) Comprehensive overview of the system’s dynamic behavior over a wide range of excitation frequencies (90 to 420 Hz) and driving forces (3 to 20 mN). The color represents periodic number of Poincaré section of 𝑥1. (C) Basin of attraction map at 236 Hz. For a fixed excitation frequency of 236 Hz, an 800 by 800 grid of initial conditions in the (𝑥1, 𝑣1) plane is integrated. (D) Basin of attraction at 270 Hz. In (C) and (D), each pixel is colored according to the k-means cluster (0 to 3) of the long-term limit cycle reached. The relationship between normalized moving displacement and excitation frequency is shown with varying driving forces: (E) 1 to 5 mN, (F) 11 to 20 mN, and (G) 51 to 91 mN. Probability of forward movement at different excitation frequencies: (H) 120 to 140 Hz, (I) 140 to 150 Hz, and (J) 259 to 383 Hz. Dark brown indicates 100% forward motion, and dark blue indicates 100% backward motion. These maps reveal the combination of parameters that generates forward or backward motions. Figure 4B provides a comprehensive overview of the system’s dynamic behavior over a wide range of excitation frequencies (90 to 420 Hz) and driving forces (3 to 20 mN). In this figure, the color scale represents the periodic number of Poincaré section of 𝑥1. At lower (<200 Hz) frequencies, the periodic number remains low, indicating relatively regular periodic motion. As the frequency and driving force increase, the color gradually shifts toward yellow, reflecting complex nonlinear behaviors. To further investigate the multistability of the system, we map the basins of attraction at selected frequencies (e.g., 236 and 270 Hz) in the initial condition space (𝑥1 versus 𝑣1). Each point in these maps is color-coded according to the attractor it converges upon. At 236 Hz, defined basin boundaries demonstrate the multistable nature of the system (clusters 0 to 3) (Fig. 4C). Figure S8 shows the attractor structures corresponding to these different clusters, demonstrating stable and different limit cycles. Figure S9 presents the fast Fourier transform (FFT) spectra of these attractor clusters. Each cluster exhibits a single dominant frequency peak at 236 Hz. The results indicate that at 236 Hz, the system exhibits multistability, with coexisting forward and backward period-1 limit cycles. Depending on the initial conditions, the system settles into different stable cycles. Similar to 236 Hz, the attractor domains at 270 Hz also display multistable nature of the system (Fig. 4D). Numerical analysis identifies four distinct limit cycles, but all correspond to forward locomotion, resulting in robust directional consistency (fig. S10). The FFT spectra corresponding to these attractors show a single dominant frequency peak (fig. S11). In addition, the phase portraits and Poincaré sections at 383 Hz are shown in fig. S12, displaying nonrepeating and complex trajectories and nonperiodic point clouds, confirming the deterministic instability in this frequency. The above results indicate that the system has rich nonlinear dynamic behaviors. To evaluate the system’s robustness, perturbations are introduced, including variations in spring and damping coefficients (±5%), amplitude (±2%), phase offset, initial velocity (±5%), and initial position (±5%). The success rate maps for forward and backward locomotion under these conditions are shown in fig. S13 (A and B). Although the effective operating regions become smaller with these changes (fig. S13C), the parameter windows with high success rates are still maintained. The above numerical analysis results demonstrate that the designed insect-scale robots exhibit diverse nonlinear dynamic behaviors, and these behaviors are different under different excitation frequency or driving force. Furthermore, the effects of the driving force, friction coefficient, position of COM, and excitation frequency on movement behaviors are explored in detail, as shown in Fig. 4 (E to J). For a fixed friction coefficient (μ = 0.4) and COM position (at front end), when the initial force is 1 mN, the motion behavior is irregular in the range of 50 to 600 Hz, as illustrated in fig. S14A. As the force increases to 5 mN, the symmetry-breaking phenomenon is observed (fig. S14B). However, at lower forces (<7 mN), the motion behaviors remain random, as shown in Fig. 4E. When the force is between ~10 and 40 mN, a boundary with distinct dynamic behaviors is observed. Similar to the experimental results (Fig. 2B and 3B), forward movements are collected at high frequency (~250 to 400 Hz). Backward movements are primarily observed at low frequency (~150 Hz) (Fig. 4F). When the force exceeds 50 mN, we cannot find the backward movement regardless of the applied excitation frequency (Fig. 4G). This result indicates that the forward and backward can be found simultaneously in different frequencies, only when the driving force is not large (<21 mN). In addition, the forward movement is observed at a higher frequency and the backward movement at a lower frequency. In this work, the driving force is provided by the electromagnetic actuator. The relevant parameters used in this model are listed in table S6. We plot a diagram to explore the dependency between normalized moving displacement and friction coefficients. In this situation, the driving force (21 mN) and COM position (at the front end) are constant, and the friction coefficients are adjusted from 0.3 to 1 (which are commonly used values). As shown in fig. S14C, the positive displacements are found at high excitation frequency (~300 Hz), while negative displacements are observed at low frequency (~150 Hz). The results demonstrate that the developed robot can achieve directional movement on different substrates with different friction coefficients by adjusting excitation frequency. The relevant parameters used in this model are listed in table S7. To investigate the effect of COM position on locomotor behavior, a revised model is established (table S8). In this model, the position of the COM is shifted from the front end to the rear end by adding an external weight to the rear end (the method is the same as Fig. 3F). In addition, the total mass of the system is unchanged. The normalized moving displacement is a function of excitation frequencies. As shown in fig. S14D, we plot the normalized moving displacement at different excitation frequencies when the COM position is set at the rear end. We find that the forward motion is in low-frequency regions, whereas the backward motion is in high-frequency regions. The robots move toward the direction of the COM located at higher frequencies and move in the opposite direction at lower frequencies. This is also strengthened by the experimental results (Fig. 2C). To investigate the operating frequency region of forward or backward motion, we calculate the proportion of positive displacement (forward motion) at different frequencies, where the driving force and friction coefficient are variable (force: 1 to 21 mN, friction coefficient: 0.3 to 0.9), as shown in Fig. 4 (H to J). In these figures, dark brown represents 100% forward motion, while dark blue represents 100% backward motion. As shown in Fig. 4H, forward and backward occur randomly between 120 and 140 Hz, highly sensitive to the combination of driving forces and friction coefficients. Small changes in the friction coefficient and driving force result in different motions. In contrast, at frequencies between 140 and 150 Hz, when the driving force is above 15 mN, the robot continues to exhibit backward motion regardless of the friction coefficients (Fig. 4I). Furthermore, forward motion is observed when the driving force exceeded 10 mN between frequencies 259 and 383 Hz (Fig. 4J). Note that the proposed Eq. 1 is a simplified one-dimensional dynamic model. It assumes that the robot does not leave the substrate when moving. However, from the perspective of energy dissipation, when the robot legs leave the substrate (as demonstrated in Fig. 3G), the ground normal reaction force and friction disappeared. In this case, the electromagnetic force will be more effectively converted into the robot’s kinetic energy, allowing the robot to move faster. Taking this into account, we optimize the simplified nonlinear dynamic model to analyze both horizontal and vertical forces at the same time (note S7 and fig. S15, A and B). Based on this specific dynamic model, the time-domain response diagrams of the maximum forward and backward speeds are calculated. The maximum forward velocity (~17.1 BL s−1) (fig. S15C) and backward velocity (~7.6 BL s−1) (fig. S15D) occur at 334 and 282 Hz, respectively, aligning well with experimental results, where the maximum forward speed of ~19.5 BL s−1 occurs at 330 Hz, and the fastest backward speed of ~9.5 BL s−1 occurs at 290 Hz (fig. S15E). The model uses a robot body length of 12 mm, consistent with the experimental prototype size. Enhanced terrain adaptability in tethered insect-scale robots To address the challenges of adapting to complex terrains, natural insects evolve intelligent locomotion strategies. For example, cockroaches can compress their bodies up to 60% in 0.3 to 0.8 s to navigate through narrow spaces (28). Depending on a unique leg structure, desert beetles can travel through loose sand without getting stuck (29). Benefiting from aerodynamic control, ants can safely land without getting hurt after falling from high places (30, 31). Therefore, in this section, we further demonstrate the terrain adaptability of the tethered insect-scale robot through the probability of successful exploration in different terrains. Given that frequency modulation is relatively simple, in the subsequent sections, we only change the frequency value to adjust the motion behaviors of the developed robot. (i) Traveling through granular media in a flipped posture. A typical characteristic of granular media is a lower penetration resistance, resulting in flowing and deforming in response to robot interactions. To achieve good traversability, it is beneficial to decrease the penetration rate by increasing the contact area with the ground. This reduction in penetration rate facilitates the efficient movement of robots across granular media. As previously described, the robot can obtain a flipped posture when the first-order resonant frequency is applied (movie S9). Compared to an upright posture, the flipped posture provides a larger ground contact area, preventing the robots from sinking and sticking on slightly concave and convex surfaces. The best speed of the robot after flipping over is achieved at ~340 Hz (fig. S16A). As shown in Fig. 5A, the robot successfully traverses through various types of sandy terrain, with granularity ranging from ~0.1 to 2 mm (movie S10). The best moving speed of ~7 BL s−1 is achieved when the granularity of sand is ~1 to 2 mm (fig. S16B). Inspired by the study by Liu et al. (32), which uses an eccentric motor to drive the robot and analyze its propulsion using a spring-dashpot-slider model, we develop a dynamic model to analyze our robot’s propulsion on granular media. The detailed information can be observed in fig. S16C and note S7. Briefly, propulsion arises from alternating jammed and fluidized states of the substrate when a robot contacts the granular media. It is worth noting that the current demonstration of the robot traversing over granular media only works for a flipped posture. Enhanced terrain adaptability of insect-scale robots. (A) Schematic illustration and optical photo showing the robot traveling through granular media in a flipped posture. (B) Schematic illustration and optical photo showing the robot navigation over a narrow pipe with obstacles in a lateral posture. (C) Schematic illustration and optical photo showing the robot retracing the starting point in a backward posture. (D) Optical photo showing a robot safety landing after dropping from a height of 1200 mm (100 BL). (E) Relationship between safe landing and the dropping height and the excitation frequency. (F) Comparison of safe landing capability among insect-scale robots, arthropods, and mammals. (G) Moving speeds of robots and friction coefficients on different substrates. (H) Comparison of our insect-scale robot to literature-reported micro or small robots. The compared capabilities include forward movement, backward movement, lateral movement, turning, safe landing capability, and moving granular media. (ii) Navigating over narrow spaces with obstacles in a lateral posture. In confined environments where space is too narrow to turn, insects typically exhibit two instinctive behaviors when encountering obstacles: squeezing through the obstacles or retreating to their starting points. Similarly, a robot with multidirectional locomotion can adjust its locomotion posture to enhance its traversability in special terrains. In our case, the robot navigates over a narrow pipe (length: ~140 mm, width: ~10 mm) containing scattered rocks by switching to a lateral posture. Compared to moving in forward/backward postures, the lateral posture reduces the traversal area, enabling the robot to navigate more efficiently. Figure 5B illustrates the movement process of the robot traveling through the narrow pipe: Initially, the robot moves in a backward posture along the pipe wall until it encounters obstacles. Then, it turns right to obtain a lateral posture. Last, the robot quickly passes through the rock pile in a lateral posture within 1.64 s (traversal velocity: ~7 BL s−1) (movie S11). (iii) Retracing to the starting points in a backward posture. In scenarios where the exploration path is entirely obstructed and the space is too limited for turning, the best strategy for the robot is to retrace its steps to the starting point to recover its movement capability (Fig. 5C). The robot initially moves forward until it encounters the rocky area. Despite multiple attempts, it fails to traverse the rocky area. Thus, the robot retreats to its original point in a backward posture. This backward maneuver not only enables the robot to return to familiar terrain but also prepares it for more effective navigation in subsequent attempts. (iv) Safe landing after falling from a height. Falling is a common and often unavoidable event for insect-scale robots. After drops from heights, most robots topple on impact or lose tunable locomotion even when their structures remain mostly intact. Therefore, the critical challenge of robots is not only to withstand the mechanical impact upon landing but also to immediately recover mobility without external intervention, and we define such a phenomenon as “safe landing” in this work. Previous studies demonstrate impressive robustness under extreme impact conditions. For example, Li et al. (33) report a robot that survives collisions at speeds up to 30 m/s by absorbing energy through body deformation. Similarly, Wu et al. (34) show a robot that withstands impacts exceeding 1500 times its weight. However, these studies do not assess whether the robots could resume motion after a fall. To address this challenge, some robots are equipped with additional deployable shells. When these shells press against the ground, the resulting reaction forces realign the robot’s body, enabling it to recover an upright posture after a fall (35). While effective, these systems often require extra actuators or bulky components, which are unsuitable for insect-scale designs. Fortunately, our robot is able to land safely and immediately keep moving forward without additional components when dropped from high heights. The main reason is that the deformable body and lightweight properties of our insect-scale robot can effectively dissipate energy during the landing collisions (33, 36, 37). The energy dissipation analysis can be found in note S8. As illustrated in Fig. 5D and movie S12, the robot exhibits safe landing capabilities even at a dropping height of 100 BL (~1200 mm). A series of experiments is conducted to investigate the factors influencing safe landing capability. Figure 5E illustrates the probability of a successful safe landing as a function of dropping height. The maximum safe landing percentage (>75%) is achieved when the dropping height is 25 BL and the excitation frequency is ~300 to 400 Hz. Although this capability is commonly observed in natural insects, it is challenging to be realized in insect-scale robots (Fig. 5F and table S9) (25, 30, 35, 38–47). Moreover, to investigate the adaptability on different surfaces, moving speeds and friction coefficients on different substrates are tested, including polymethyl methacrylate, glass, paper, and sandpaper (Fig. 5G). After appropriate parameter adjustments, the robot exhibits consistent forward locomotion across varying surface conditions. The best moving speed is achieved on sandpaper with a granularity of Sp-p10,000. The capabilities of climbing slopes and carrying payloads (fig. S17) are explored. In general, compared to literature-reported micro or small robots, our insect-scale robot demonstrates a well-balanced performance between maneuverability and terrain adaptability (Fig. 5H and table S10) (3, 11, 21, 24, 48–51). Tethered insect-scale robot completing tasks on cross-domain terrains The obtained insect-scale robot based on nonlinear dynamic behaviors exhibits good maneuverability and terrain adaptability. It is capable of completing various complex tasks by the integration of different movement postures (Fig. 6A). For example, Fig. 6B illustrates the robot’s ability to navigate through a narrow maze, where the path is too confined to allow turning, by combining forward, backward, and lateral movement postures (fig. S18 and movie S13). Figure 6C displays a type of potential application of the robots, i.e., inspecting and unclogging a blocked pipe using two different motion postures (upright posture and flipped posture) (movie S14). Tethered insect-scale robot completing tasks on cross-domain terrains. (A) Various movement postures used in the demonstration. (B) Optical photo showing robot traveling through a narrow maze. (C) Optical photo showing robot unclogging in a flipped posture. (D) Optical image showing a robot dropping from a high height and reaching four points on a planar. (E) Optical photo showing the robot completing challenging tasks relying on multiple movement postures. Moreover, the robot is deployed in various hypothetical scenarios. First, the robot is placed on a relatively simple terrain with only a high place and flat ground. In this case, the robot drops from a high place (height: 20 BL) and successfully lands on the solid substrate. Then, the robot quickly reaches four cardinal directions on the plane in the integration of multiple movement postures, involving forward, backward, lateral, forward-left turn, and backward-left turn postures (Fig. 6D and movie S15). Then, a challenging cross-domain terrain is proposed, involving a narrow and partially obstructed pipe, a high platform with cutoffs, a granular area, and solid ground. As shown in Fig. 6E, the robot moves in a forward posture along the narrow pipe. Subsequently, the robot crosses the rocky area in a lateral posture, taking about 2.6 s. Then, the robot drops from the high place. After 3.3 s, the robot lands safely on the ground and continually moves forward. However, its moving path is blocked by sand. After several unsuccessful attempts to cross the sand in an upright posture, the robot actively flips when the first-order resonant frequency is applied. Then, the robot successfully traverses the granular media in a flipped posture. It is worth noting that the robot completes all the above tasks in only 20 s, exhibiting excellent maneuverability and terrain adaptability (movie S16). Practicality evaluation of untethered prototype The untethered prototype is twice the size of the tethered prototype, dictated by the dimensions of the designed circuit board (fig. S19A and table S11). As demonstrated earlier, the diverse behaviors of the insect-scale robot can be modulated by adjusting the excitation frequency. To enable controllable movement of the untethered prototype, we develop a miniaturized circuit board integrating power and control circuit units (Fig. 7, A and B). The circuit board is powered by a battery (150 mAh, 3.7 V, 3570 mg) and includes two voltage regulators, a microcontroller unit (MCU), a Bluetooth module, and an H-bridge driver. This setup allows for the output of a single driving channel with variable frequencies, controlled wirelessly by a host computer. Detailed circuit design information is available in fig. S19B and Materials and Methods. The mechanism by which the untethered prototype changes its movement direction is the same as that of the tethered prototype: When a higher excitation frequency is applied, the robot moves toward the side where its COM is concentrated, while at a lower excitation frequency, it moves toward the opposite direction. Diverse movement demonstration of the untethered prototype. (A) Block diagram illustrating the designed power and control circuit units. (B) Top view of the power and control circuit board. (C) Untethered prototype moves toward the COM located, when a higher excitation frequency is applied. (D) Untethered prototype moves in the opposite direction, when a lower excitation frequency is applied. (E) Optical photos showing clockwise and (F) counterclockwise rotations of the untethered prototype. Optical photos showing the untethered prototype (G) moves laterally, (H) climbs slope, and (I) travels through granular media. (J) Untethered prototype turns left and right in an outdoor environment. MCU, microcontroller unit. As shown in Fig. 7C, the battery placed on the rear leg causes a shift in the COM toward the rear end. This configuration enables the prototype to move backward at a high excitation frequency (125 Hz) and forward (Fig. 7D) at a low excitation frequency (50 Hz). Moreover, movie S17 demonstrates that the untethered prototype is able to perform both clockwise and counterclockwise rotations with a small turning radius under different excitation frequencies. Specifically, the prototype rotates clockwise at 70 Hz, whereas a higher excitation frequency (170 Hz) induces a counterclockwise rotation. Notably, a 90° clockwise turn requires ~2.5 s (Fig. 7E), while a 90° counterclockwise turn takes ~5.6 s (Fig. 7F). In addition, the backward-left and backward-right turning motions of the untethered prototype are illustrated in fig. S19 (C and D) and movie S18. Similar to the locomotion behavior of the tethered prototype, the untethered prototype is also able to move laterally (Fig. 7G and movie S19), climb slope (Fig. 7H and movie S20), and traverse granular media (Fig. 7I and movie S21) by changing its locomotion posture. To assess the robot’s locomotion performance beyond laboratory conditions, we extend the experiment to the outdoor environment. The robot successfully executes left and right turns on concrete floors in 14 s to form an “S” shape trajectory, as shown in Fig. 7J and movie S22. This further validates the reliability of our modulation strategies based on the nonlinear dynamic behavior and highlights the practical potential of the proposed insect-scale robots. DISCUSSION In this work, we propose insect-scale robotics based on nonlinear dynamic behaviors to address the inherent tension between the simplicity of design and the preservation of functional versatility. By introducing asymmetries into a deformable hexagonal frame, the insect-scale robot is developed, demonstrating high maneuverability and terrain adaptability, similar to that of natural insects. These capabilities highlight its potential for applications in complex terrains, such as urban disaster rescue and pipeline inspection. Although our robots demonstrate versatile mobility enabled by the nonlinear dynamic-based modulation strategy, several limitations still exist. First, the current prototype lacks the capability to climb vertical surfaces or overcome high obstacles. A promising future direction would be to enhance climbing performance, potentially by integrating reversible adhesive materials or electrostatic footpads capable of rapidly switching between adhesive and nonadhesive states via electrical stimulation. Second, autonomously achieving reversible transitions between upright and flipped postures remains challenging. Now, the robot relies on passive interactions with the environment, such as falling from a height, to passively induce reorientation and regain its upright posture (fig. S20 and movie S23). Thus, another key future objective is developing an active and autonomous self-righting mechanism (e.g., tumbler-like structure), enabling the robot to reliably recover from inverted positions without relying on environmental assistance. Third, most of the complex tasks are demonstrated using the tethered prototype. Therefore, future efforts will focus on extending these capabilities to untethered platforms. A promising solution involves assembling multiple untethered robots into modular system. Hence, the modular robots not only can keep the versatility of each individual unit but also can coordinate to accomplish complex tasks. Our way of achieving versatility through a simple and efficient modulation strategy is also helpful to reduce the general complexity of modular system. MATERIALS AND METHODS Fabrication of insect-scale robots The material of the hexagonal frame of the robot is polyethylene terephthalate with a thickness of 0.15. The electromagnetic actuator consists of a magnet (NdFeB, N52) and a planar coil. The radius of the magnet is 2.5 mm, and the height is 3 mm. All planar coils are made of copper wire with a diameter of 0.07 mm. The inner diameter of the planar coil is 1.8 mm, and the outer diameter is 5 mm. In addition, there is no gap between each coil. The distance between the coil and the magnet is ~2 mm. Power and control electronics To achieve controlled locomotion of the untethered insect-scale robots, we develop a miniaturized double-sided board with the power unit on one side and the control circuit on the other. The input voltage from the lithium battery first passes through two regulators. Regulator 1 adjusts the input voltage from 3.7 to 3.3 V to ensure effective operation of the MCU, while regulator 2 provides a 5-V output to power the H-bridge driver. Wireless commands generated by a computer are transmitted to the MCU via wireless fidelity. Variable excitation frequencies are obtained from the H-bridge and applied to the robot. Power consumption and COT calculation Energy efficiency is characterized by the COT, defined asCOT=p/mgv(2)where p, m, and v are the power consumption, mass, and running speed of the robot, respectively. The power consumption of an untethered prototype, including the control board and an electromagnetic actuator, is measured to be ~0.25 W (fig. S21A). For an untethered prototype (9.72 g), which works at 80 Hz, its maximum moving speed is 4 mm s−1 (fig. S21B). Therefore, the COT of the untethered prototype is ~656. When the battery capacity is 150 mhA, the measured maximum duration of the untethered prototype can reach ~7.5 min (fig. S21C). Characterization The magnetic flux of the magnet is measured by a commercial meter (HT20, Hengtong Magnetics). The displacement is measured by a laser Doppler velocimeter. The voltage and excitation frequency signals are generated by a function generator (Keysight 33210A). The waveform used in this work is a square wave. A high-speed camera (Phantom VEO 710S) captures the motion of the robot at a sampling frequency of 5000 fps. The power consumption of a robot is measured by PA310 high-accuracy digital power meter (Zhiyuan Electronics, China).