Synthesis of mass-spring networks from high-level code descriptions

TL;DR

This paper addresses the challenge of designing nonlinear elastic systems that perform specified tasks by encoding behavior in a Mechanical Description Language (MDL) and automatically compiling it into a mass-spring network with embodied intelligence. The authors introduce a three-stage mechanical compiler that translates MDL into a digital gate/netlist, maps it to a network of bistable masses with variable-stiffness couplings, and then lays out the physical device under geometric constraints. They demonstrate two functional designs—a maze-navigating robot and a numerical combination lock—generated from MDL and validated via numerical simulations, with the added capability of translating natural-language descriptions into MDL code using Large Language Models. The approach enables on-demand, programmable intelligent devices and offers extensibility to new sensors, actuators, and three-dimensional geometries, highlighting a pathway toward practical mechanical computation.

Abstract

Structural nonlinearity can be harnessed to program complex functionalities in robotic devices. However, it remains a challenge to design nonlinear systems that will accomplish a specific, desired task. The responses that we typically describe as intelligent -- such a robot navigating a maze -- require a large number of degrees of freedom and cannot be captured by traditional optimization objective functions. In this work, we explore a code-based synthesis approach to design mass-spring systems with embodied intelligence. The approach starts from a source code, written in a \emph{mechanical description language}, that details the system boundary, sensor and actuator locations, and desired behavior. A synthesizer software then automatically generates a mass-spring network that performs the described function from the source code description. We exemplify this methodology by designing mass-spring systems realizing a maze-navigating robot and a programmable lock. Remarkably, mechanical description languages can be combined with large-language models, to translate a natural-language description of a task into a functional device.

Figures (6)

• Figure 1: Automated mechanical synthesis design.a, Design flow for mechanical synthesis. The designer specifies the desired behavior of the system using a Mechanical Description Language (MDL). This description is interpreted by a mechanical compiler which generates the design accordingly. b, The mechanical compiler first converts the MDL code into a gate-level netlist, then maps the list to its physical equivalent using elements from the library made of bistable-masses and variable-stiffness couplings, and finally assigns the spatial coordinates while satisfying geometric constraints and prescribed input/output locations.
• Figure 2: Mechanical description language and generated mass-spring networka The keyword mechanicalmodule is used to identify and name a component---here named mazerobot in correspondence with example I. b The shape (boundary) of the element is defined as a rectangle, composed of a set of straight lines---identified by the keyword line. The lines are specified by a name and two integer endpoints, and listed in clockwise order---enabling the algorithm to identify the inside/outside of the robot. c The location, and type of the sensors and actuators is specified through sensor and actuator elements. The location is specified as a coordinate along a specific wall, while the value field maps sensor measurements and actuator actions to corresponding binary values. For the purpose of this example, we include only one type of sensor and one type of actuator, but in principle large libraries could be constructed. d The next block of the code defines the logical behavior of the system, using the Verilog syntax. e An example synthesized mass-spring network generated from the MDL code using the mechanical compiler. The circles represent the masses in the network (with the color corresponding to the biasing force). The lines represent interacting potentials as described in Fig. \ref{['fig:LibraryElements']}, with the red bars indicating negative interaction potentials.
• Figure 3: Mechanical logic building blocks.a, The basic building block is a mass, subject to a potential that can be switched between single and double well configurations. The power clock (blue mass) modulates the potential from bistable to monostable, enabling state switching. b Pairs of masses are connected through nonlinear potentials, whose energy form depends on the displacement of each mass ($x_i$, $x_j$) and the power clock ($x_p$). Because the power clock is externally prescribed, the coupling potentials effectively behave as a time-dependent, pairwise interaction. Depending on the specific energy form, this effective interaction can be linear or nonlinear. c Top left: The universal NOR gate design, implemented by in a four-mass system using negative effective springs and a local biasing potential. Bottom left: Numerical simulation of the time-evolution of the on-site potentials acting on the masses. When the coupling of the inputs and the intermediate mass are strong, the inputs are subject to a double well potential while the intermediate mass is monostable---allowing it to follow the inputs. As the information propagates one site forward, the coupling weakens between the input and intermediate masses while strengthening between the intermediate and output masses. This cycle is then repeated. Right: The displacement amplitude of the inputs and outputs over the power clock cycles, exhibiting a NOR gate response. The positive (negative) displacements shows that the corresponding mass takes $1$ ($0$) logical state. d When multiple inputs are connected to a single output, we use a branching mechanism constructed from buffer gates. e A NOT gate is similar to a buffer, but it inverts the input signal through a negative effective coupling. The graph shows a numerical simulation of the input and output displacements as the power clock cycles. f, Clock circuit driving a D-latch. We use a loop of masses of length $N$ to create a periodic clock signal. The D-latch consists of four masses: Two inputs (clock $C$ and data $D$), an intermediate mass, and an output. The clock input $C$ is connected to the intermediate mass via a nonlinear interaction, causing the potential to switch from monostable to bi-stable when the clock ticks. The data input $D$ biases the intermediate mass towards $0$ or $1$, causing a specific value to be latched when the single well-double well transition is triggered by a clock tick. Longer effective clocks can be constructed by connecting multiple, co-prime loops through an AND gate. The plot presents a numerical simulation of a clock-latch tandem. The latch input $D$ (orange) is copied into the output $Q$ (green) when the clock $C$ input ticks (blue).
• Figure 4: Numerical simulation of a maze-following robot.a Path taken by a robot generated from a MDL description. The mass-spring model (Fig. \ref{['fig:PLace&Route']}e) is simulated via time-integration, subject to external sensor forces that depend on the position within the maze. b The robot response is governed by a finite-state machine (FSM), with four states corresponding to the four possible directions of motion left, right, up and down. These states are mapped to actuator outputs of $00$, $10$, $01$ and $11$ respectively. The FSM updates its direction based on the sensor inputs, that detect the walls around the robot. c, The initial state is left ($00$). Sensor inputs indicate that all the directions are blocked except right. Therefore, The system switches to right motion, updating its outputs to $10$ after receiving the clock signal. d, In step 19, the FSM has a directional state of up ($01$) and the sensor inputs indicate that right and front are blocked. Following the FSM update rule, the machine turns left, ending in state $00$. e, In step 37, the directional state of the machine is right ($10$). Based on the sensor inputs the FSM updates its state to down motion by switching to the $11$ state.
• Figure 5: Synthesis of one-time code lock.a, The synthesized mass-spring network embodied within a defined boundary. The system has inputs representing the four-digit binary passcode, with masses representing key and action, as well as a door output mass. b, The system operates with two primary states of unlocked ($0$) and locked ($1$). When locked by applying the passcode while both key and action are set to state $1$, the system falls into one of sixteen sub-states of locked, corresponding to all possible passcode combinations. To open the lock, the registered passcode together with the action should be pressed. c, Locking process. Top: We apply a force on the action and key masses to set the state $1$ while applying a desired passcode ($1001$). Bottom: Initially the system is at unlock state, with door mass being at the negative stable state ($0$), after the clock signal, the passcode is registered and the door moves to positive stable state ($1$). Then, the system is locked. d, Unlocking attempt with an incorrect passcode. We set the action key to positive stable state ($1$), along with an incorrect passcode ($0011$). The output stays in the positive stable state, and the system remains locked. e, Successful unlocking. Using the correct passcode ($1001$) together with forcing action input to ($0$), causes the door to change state from $1$ to $0$, and the system becomes unlocked.