You are on page 1of 8

Computational Design of Walking Automata

Gaurav Bharaj1∗ Stelian Coros2 Bernhard Thomaszewski3 James Tompkin1 Bernd Bickel4 Hanspeter Pfister1
Harvard SEAS 2 Carnegie Mellon University 3
Disney Research Zürich 4 IST Austria

Figure 1: Two walking automata designed with our system and the corresponding fabricated prototypes.

Abstract Modern rapid and additive manufacturing techniques expand the
scope of mechanical design and construction. Motivated by this
Creating mechanical automata that can walk in stable and pleasing tremendous progress, the graphics community has started to em-
manners is a challenging task that requires both skill and expertise. brace the challenge of translating digital characters into physical
We propose to use computational design to offset the technical artifacts [Zhu et al. 2012; Coros et al. 2013; Ceylan et al. 2013;
difficulties of this process. A simple drag-and-drop interface allows Thomaszewski et al. 2014]. These methods can create virtual charac-
casual users to create personalized walking toys from a library of ters with a broad range of complex motions, such as an in-air mimic
pre-defined template mechanisms. Provided with this input, our of a walking motion, but their fabricated physical counterparts have
method leverages physical simulation and evolutionary optimization yet to demonstrate an ability to actually walk stably. This is because
to refine the mechanical designs such that the resulting toys are neither the mechanics nor the geometry of stable locomotion are con-
able to walk. The optimization process is guided by an intuitive set sidered during the design. Considering these constraints would allow
of objectives that measure the quality of the walking motions. We users to explore the space of feasible linkages for stable walking,
demonstrate our approach on a set of simulated mechanical toys with and so to more easily design walking automata.
different numbers of legs and various distinct gaits. Two fabricated
prototypes showcase the feasibility of our designs. Capitalizing on this work, our method allows users to intuitively
create unique automata designs that walk stably once fabricated.
CR Categories: I.3.7 [Computer Graphics]: Three-Dimensional First, we learn a space of linkage configurations which are likely to
Graphics And Realism—Animation lead to stable walking. Then, the user designs the automata by plac-
ing 2–4 linkage templates onto a body at arbitrary positions. From
this initialization, we optimize the overall mechanical structures of
Keywords: Mechanical characters, animation, fabrication the automata, allowing them to automatically discover how to walk
with the same intuitive set of objective functions. This approach
1 Introduction integrates physics-based simulation of mechanical assemblies with
an evolutionary optimization algorithm that is able to explore the
Stories of walking automata date back to ancient Greece, with the complex design space of these structures. We demonstrate our ap-
Renaissance reviving the tradition and providing prominent histori- proach with simulated designs, and we validate our simulations by
cal examples: da Vinci’s mechanical Lion, from the 15th century, fabricating two very different walking designs for a dog and a crab.
delighted audiences upon its rebuilding in 2009. Although toy stores
abound with mass-produced walking automata, their design remains 2 Related Work
challenging. It is arguably for this reason that many commercial
walking automata are based on a small number of template mech-
anisms such as the famous Klann or Jansen linkages. But, even Character Animation One major inspiration is Sims’ [1994] pi-
in this restricted setting, the design problem is still far from obvi- oneering work on virtual creatures that discover how to locomote.
ous: in addition to satisfying kinematic requirements on individual Sims uses genetic algorithms — a class of evolutionary optimization
mechanisms (i.e., producing desired end-effector motion), the de- methods inspired by natural selection — to discover the structure
signer must carefully select parameters for all mechanisms, includ- and morphology of virtual creatures, and to discover time-varying
ing timing and inertia, to obtain stable walking motion. This task is actuation forces that lead to crawling or jumping motions. This
challenging for experts, and so is often beyond the capabilities of general approach has been successfully adapted to a large range of
average users. additional animation domains, such as realistic control of swimming
[Tan et al. 2011], muscle-driven biped simulation [Geijtenbeek et al.
∗ Email: 2013], gait discovery for quadrupeds [Lee et al. 2013], or learning
bicycle stunts [Tan et al. 2014].
Translating virtual walk simulations into the real world is non-trivial.
For humans and animals, hundreds of muscles have to act in unison
via a central nervous system for stable and efficient gaits. In a robot,
the orchestration of actuators as muscles requires many sensors and
a complex controller. In this context, Sims-like virtual characters are
difficult to fabricate as, even if one could find physical actuators and

yet. and especially for the simple ters pk . 2013]. timing parameters pt per timing mechanism are used to control the relative phase between different linkages by directly specifying the Walking Motions Control in Animation and Robotics A variety phase profile functions of the virtual actuators. The fabrication process characters using elastic materials such that their deformation under is manual. we fabricate the optimal automata. 2012. Lee et al. This highlights the need for automated methods. unless we used a large number of legs (i. [2013]. and sometimes interesting gaits to . to be driven by a single servo motor with speed control. 2a. Starting from the input configuration created by the user with our To our knowledge. This data-driven approach is a reparameterization 3D printable representations of virtual characters with joints [Bächer equivalent to learning and navigating a manifold on the original et al. Each linkage is assumed one motor per limb. recent works present custom-shaped objects that can Not all linkage parameterizations are valid configurations which will fly [Umetani et al. Initially. yet are still able to perform walking motions. 2014]. with metal bolts for joints. 2014]. optimize the design to walk. [2013] note that even if the motion of a mechanical and servo motors for drive (§6). linkage belongs to a set of linkage classes. Coros et al. with each parameterizable which allows for stable. Some methods aim to bring tion. body in a rigid-body physical simulation (Supplemental appendix A). Physical prototypes are manu- factured using a laser cutter and plywood. character at first glance resembles walking. joints for all virtual motors. with all kinematic parameters and motor controls output the influence of external forces can be controlled. These linkages can of advanced control methods have been proposed for antropomor. complex control strategies require complicated mechanics. 2013.. we must automatically designing automata that can walk stably. such as Klann or Jansen linkages. we employ a simpler in nature. Our designs are significantly body (Fig. stand on their own [Prévost et al. Our initial database of pre-configured linkages was created using Such methods have been applied to sophisticated legged robots to the method of Coros et al. and so to speed up this optimiza- or spin stably [Bacher et al. Each from robotics. signed automatically [Coros et al. we provide a simple drag-and-drop design tool to the user. formed from high- 3 Overview dimensional parameterizations. Rather than attempting the very difficult task of explicitly applying locomotion control knowledge Each of our automatons has a body with 2–4 linkages attached. and the StarlETH robot is well beyond the complexity and who chooses and places linkage classes at arbitrary positions on the cost of our target of automata as toys. 2013]. In initial experi- ments. encode the location of each pin joint.e. Thomaszewski et al. which is unlikely to walk. and ac. our work is first to investigate the challenge of simple interface. 2014]. be standard mechanisms. 2013. of interesting (non-walking) motions [Coros et al. this does not mean that the character would actually walk if fabricated. 2014]. Calı̀ et al. our work is much closer to recent work in computational (§4) to optimize the linkage instance parameters with respect to the design than to the general field of robotics. the resulting cost would exceed what by pin joints and timings (Fig. parameters of the automata — the mechanical linkage structures — ficulty of design and manufacturing problems by creating tools change. and so define the automata that we consider: for toy and educational use. they have mechanical configuration of the automata. Right: The user analyzes the final walk optimized by our system. we were not able to create any automata in this way that 4 Quantifying Walking were capable of walking stably. 2013. to design mechanical toys capable high-dimensional space (§5). For instance. 2012]. Walking is complex. The simulation measures the quality of walking motions as the Computational Design and Fabrication This field reduces the dif. or to manufacture physical Finally. We show two initial linkage generate controllers [Gehring et al. Thomaszewski et al. move or lead to smooth motion. or de- phic physically-simulated humans [Geijtenbeek et al. of locomotion controllers [Gehring et al. 2013. 2011]. However. 2010] and animals [Wampler and Popović 2009. Offline Function Learning Metric Weights Distance Time Smoothness Effort Shape-Regularizer Automata Sampling Physical Simulation Drag-n-drop linkages Metric Cost Preconfigured Linkage Interactive Optimized Database Automata Creation Stochastic Optimization Walking Automata Figure 2: Overview: Left: We learn valid linkage configuration functions from a database of prefigured linkages. It also previews the physical prototype output. efficient. and supplemental video). or to increase the agility configurations in supplemental appendix B. Coros et al. Mid-left: The user designs an automata with a drag-and-drop interface. Following this. tuators. The 12-16 kinematic parame- is acceptable for most applications. sensors. which forego or reduce the need for expert domain knowledge. instead we describe an intuitive set of sub-objectives. Mid-right: Stochastic optimization using the valid linkage functions and physical simulation allows the automata to walk. directly by the optimization process. Ceylan et al. and it is now possible to create configurations. and specifies a metric by changing weights. We assume arbitrary kinematic configurations. 3). 2014]. requiring muscles (or motors and sensors) to act in unison for stable and efficient gaits. stochastic genetic algorithm (CMA) and a walking objective function As such. and no high-level controller. hexapod). no sensors. so the four they can walk successfully once fabricated. we precompute within linkage space a subset of good linkage virtual characters to the real world.

The recorded simula- tion states are used to evaluate each individual sub-objective (below). we found that the optimization could not from exploring more relevant regions of the parameter space (see recover from such local minima. Therefore. can quickly Distance: We wish the mechanism to walk forwards.e. make little improvement as the spaces are not smooth. [2013]. However. even around valid configurations. [2014] analyzed the singular values of the constraint Jacobian ∂C/∂q (see Supplemental appendix A) as a continuous measure of the distance away from singular configu- rations. . and attempt to learn a function of ferences on the generalized velocity vectors q̇ at consecutive time valid kinematics parameters from which to draw more useful samples steps.. [2012]). the automaton would often perform which at least one degenerate linkage was detected. t < T ) if parts of the automaton other than the feet or linkages is unlikely to lead to a successful walk. and non-smooth due to the unilateral and in- termittent nature of the contact forces.. Therefore. The simulation proceeds until a failure mode is en. Simulation position and orientation states and derivatives. we add a term which encourages the automaton to remain upright To alleviate this problem. at the beginning and at the end of the simulation. not on the lead to degenerate mechanism that grind to a halt in mid-step or spot. construct a reparameterization of the space of linkages such that (almost) all samples generated by CMA correspond to valid linkage Smoothness: We penalize the acceleration of the center of mass configurations. To avoid these solutions altogether. Since we physically simulate. We define Supright = (T − t)2 . The acceleration a is estimated using finite dif. we introduce a simple regularizer for the to a change in the motion of its end effector (red curves). are parameters optimized Regularizer: Since the input to our system consists of a user- by our method. as this quickly increases the distance traveled. are recorded for each timestep t. To this end. small moment arms require larger motor torques and result in large internal forces acting on the me- Linkage chanical structures. To not waste computation time. Increasing during CMA. The simulation is stopped walking. or until a fixed amount of simulation time T = 30s has [Wampler and Popović 2009]. Worse. where v is a unit vector that points along the correspond to valid configurations. al. the Gaussian distribution is updated. We define the walking objective as the weighted sum: 5 Optimization Stotal = ωd Sdist + ωt Supright + The optimization problem introduced in the previous section is high- ωs Ssmooth + ωe Sef f ort + ωr Sregularizer (1) dimensional. The penalty term is defined as Ssmooth = at a. each methods are ill-suited.. as noted by Coros et. singular configurations) are pruned early on. and so we measure the vector d between the center of mass cannot move at all. nor is it an incremental solution towards to adapt its internal distribution and ultimately prevent the algorithm better walking. However. which de- fine the kinematics of the mechanical toys. For instance.1). for each explored set of parameters. kinematic parameter problem. i. the weight of this term leads to more conservative motions which we Crank posit increases the likelihood of successful walking once fabricated. where t is for validity before optimization to learn valid configurations for the time when the simulation is stopped. Our experiments confirmed these observations. arise. since we minimize the total score Stotal . This seemingly simple and efficient. before running any simulations.e. Changing the location of a pin joint (yellow arrow) created automaton. while a somersault. Consequently. until convergence. so we distribution. λ in Supplemental appendix A. eq.e. we define Sef f ort = λT λ and add it to Stotal to mini- mize the net internal forces acting throughout the mechanisms. also Xu et al. kinematic parameters pk : Sregularizer = (pk − p0k )T (pk − p0k ). and the process repeats exceed a threshold value. and define Sdist indicating that only ≈30–40% of the randomly generated samples simply as −vT d. One solution might be to first check all linkages of the automaton in isolation and reject samples for Upright: In early experiments. increasing the likelihood of mechanical failure. We would like to bottom of the body touch the ground. gradient-based During optimization. and we resort to a derivative-free stochastic automaton configuration. and then physically simulate the trix Adaptation algorithm [Hansen 2006]. Figure 3: The locations of the pin joints (yellow circles). Thomaszewski et al. we compute the mass and moment of evolutionary optimization technique based on the Covariance Ma- inertia of each rigid body part. q̇). we have direct access to Traced-Curve the magnitudes of the forces needed to satisfy the joint and motor constraints (i. desired walking direction. CMA generates parameter samples according to an internal Gaussian Some parameter settings will result in infeasible mechanisms. non-linear. as well as as possible. The negative sign is used to promote it would be desirable to rule out such degenerate configurations even walking longer distances.. the parameter spaces of mechanical linkages are highly non-linear and random sampling. an errant pin joint configuration which locks early (i. After evaluating the objective value for each sampled monitor constraint values and terminate the simulation early if they point. we turn to Gaussian Mixture Models. this strategy would not allow CMA is not successful walking. Therefore.e. Pin-Joint Effort: While infeasible mechanisms (i. for our problem this would elapsed. s = (q. Some works attempt to automaton. For example. where p0k represents the initial parameter values. this does not mean that all mechanisms are equally desirable. however. we note that linkages can be analyzed for as long as possible. improve CMA by adding derivative-based local constraint satisfiers countered. we would like to change the design as little leads to a change in the structure of the mechanical leg. we compartmentalize the valid linkage at every timestep.

and now need a structured Parameter Learning Given sampled data D. where Θm . this can be used to answer queries such as: N X M X  ln πm N (di | µk . 1978] to determine the number of categories. For a given sition and extract the eigenvectors (E m ) and eigenvalues (Am ) of . Data Preprocessing For a linkage. Sampling by Inference We have learned a GMM over the space of kinematic parameters for a linkage. 000 samples sufficed. we perform PCA on the learned C is modeled as a distribution p(C) = M cm Q m=1 πm s.e. i. m gives the mixture index and x (a bijective mapping to pk ) [Schwarz et al. 2011]. Mathematically (in log space) this is defined as: X = {C. the kinematic parameter pk between two bars. -2 We want to model the space of valid configurations with multiple Gaussians in a mixture. As this is a high-dimensional space. Σk ) (3) p(x | cm = 1) = Nm (x | Θm ) (4) i=1 m=1 Initially. 2006] to create a parametric function over valid spaces: 0. We vary these points of location Leave-p-out cross validation [Hastie et al. for a given linkage. for each step of a full motion cycle and label a sample as valid. we use a 1 multivariate Gaussian mixture model (GMM) [Bishop et al. we set cm = 1. One advantage calculate Gaussian parameters Θm = {µm . the BIC criteria. thus sampling the whole space. Next. Since all the covariance matrices are by 12 M log(|D|). symmetric positive definite. and models the kinematic parameters of the linkage configuration. p(cm = 1) = πm . we found that 6 − 12 mixtures gave us good results. Given a number of mixtures (which we -1 -0.5 M X 0 f (x) = πm N (x | µm .5 0 0. we draw ples fit the GMM model’s probability-density function.5 matrix of the mth multivariate Gaussian in the mixture. where l is a vector of lengths of the cor. and Σm is an n×n covariance -1. so we use Expectation Maximization [Bishop et al. and is given by a multivariate GMM. GMM. Σm } and πm which of this probabilistic approach is that the learned random variable best model it: we wish to maximize the likelihood of D given parameters encode the joint probability distribution P (X). It is noted samples D from the space of possible linkage configurations using that full covariance matrices Σm can lead to data over-fitting. Σm ). for each sample.5 Second Principal Component pockets of space in which valid samples exist. For example for the linkage configuration given in Figure 4. linkage. 0 ≤ πm ≤ 1 and M P m=1 πm = 1 are the mixing coefficients. we penalize Equation 3 precomputed Gaussian (Eq.. Using similar parameters. and use the one that maximized the cost. πm . the above mentioned over-fitting checks.5 m=1 -1 Here. and thereby the number of Gaussians in the index is selected.5 1 discuss discovering later on). i. with one element is ever one. this means that it is highly likely that a given unknown. in an Number of Mixtures & Over-Fitting The number of components automaton. x] (5) model-parameters. 2006] linkage parameter x belongs to the valid space and is modeled to find the maximum likelihood estimation of the random variable by the mth Gaussian (a conditional distribution). the kinematic parameters of a linkage pk are replaced by in the mixture plays an important role in determining the effective. we use the Bayesian Information Criteria Here. the following vector: ness of the learned model. first we perform eigenvalue decompo- ber of mixtures. can vary over the Another method that we employ to avoid over-fitting the data is lengths of the connecting bars. the marginal space and show only the first two principal components.1 Linkage Configuration Function Learning 2 Learning Parametric Function We observe that there are many 1. However. while too many mixture components can lead to over-fitting. when number strategy uniformly subdivides the sample space and ensures that a of Gaussians was twice as large as the number of mixtures in the sample is drawn from each division. 4). we precompute M multi- variate distributions modeled by the GMM. each pin joint location. we would like to way to sample from the GMM for use in CMA. we introduce a random latent variable First Principal Component C which assigns a linkage configuration to a Gaussian in the mixture. for our problem using 6 − 12 mixtures with full Then. which selects the corresponding mixture. This gives us the desired linkage-configuration parametric function (equation 2). inference queries for all valid m mixtures. with all others zero (1-of-M representation). the dimensionality of the 1-of-M representation of C is More intuitively. 1979]. are the multivariate Gaussian function parameters. We sum the energies we found that using |D| = 25. over multiple ps. pendix A) for pin-joints and motor-constraints. if the summed energy is below a given threshold value (we use 10−2 ). Σm ) (2) -0. This sampling experimented with only diagonal covariance matrices. K} and factorizes as P (X) = P (K|C)P (C). N (x | µm . Using these box limits. 2009]. Random variable K is a continuous and observed. This validation indicates how well the test sam- responding bars in the linkage. distribution over C is given by the mixing coefficients. This is indicated by an m-dimensional binary variable where only Figure 4: We show a sampling of the learned linkage space. Finally.5. As such. we Latin-Hypercube Sampling [McKay et al. Depending on the dimensionality of the problem. using too few mixture components can lead to inaccurate p∗k = [m. we perform this within − 4l ≤ pk ≤ 4l .t. by sequentially varying the num.. Hence. we found that six components maximized the EM energy in Equation 3. When the mth the dimensionality of C. µm is an n-dimensional mean vector. Hence. examples from three different Gaussians in the mixture (color coded).e. we simulate the motion of the corresponding covariance matrices gave us the better results while still satisfying linkage by minimizing the constraint energy (see supplemental Ap. Similar to [Chaudhuri et al.

is not CMA’s internal multivariate Gaussian is sampled giving p∗k . CMA vs. very inefficient gait with the dog essentially moving in place. we ran eight trails for Grandpa-Bot. which increases stability Since the mixture weights πm give the probability that a sample from and thus leads to a successful walking motion. we can have a different phase. we found that. their kinematic design tool did was confirmed in all examples that we considered. We validate our CMA-GMM improvements over many trials. having no notion of the both faster and produces solutions with better scores—a trend that physics or geometry of locomotion.. optimizing only for the timing parameters. left. 7a. right). leads to a such Gaussians in the mixture. Optimization Summary We precompute a database of linkage Lobster Our second example. rameters. we run CMA.. of Iterations (x Populations-size) Figure 5: Trial runs with CMA vs. Optimizing for the full set of 20 structural and 10 the linkage parameters pk are discovered using p∗k and Equation 6. linkage sampling. allowing for an asymmetric structure. we invite configurations in Statistics for all examples are listed in Table 1. can be seen in Fig. and video). While this starting configuration falls over immediately at the beginning of √ the simulation. CMA-GMM. the can quickly move between different Gaussians and can explore valid various imprecisions introduced through simplifying assumptions configuration spaces quickly and effectively. comparison of the convergence rates for CMA and CMA with GMM Our Grandpa-Bot is based on an example used by Coros et al. As a result. Then. the valid configuration space belongs to Gaussian. 14 structural and 4 velocity profile parameters for the two front and hind legs. is essential for obtaining a successful gait: as can be seen in the the mixture-index parameters m are weighted by πm and mapped. Accord. over which we learn an M -dimensional GMM. gait that alternates between swing-through and support for the outer ingly. [2013]. and manufacturing make such dynamic motions more susceptible to mization can also be interpreted as a discrete (mixture-component instabilities. accompanying video. with the two front and hind legs having the same mechanical structure. While each leg The fourth example is our Gorilla-inspired autoamata (Fig 7b). Figure 4 shows samples with random x inputs. of the smoothness and effort objective and introduced the height of the body as an additional parameter for the optimization. We seen that. Figure 5 shows a automata in simulation to further explore the behavior of our method. appearance.5  Table 1: Statistics for all examples. . Again. Then second. so is a dynamic with the center of mass shifting back and forth (see ac- drop-in replacement. i. the resulting motion was in- we generate 4 + b3 × log(N )c samples for every CMA iteration. respectively. able to walk. Iterations (hrs) Trial CMA nisms  Average CMA 34 4 20 (Purple) 450 3 Dog 40 (Blue) 500 4 Log Score  Grandpa 20 2 13 (Purple) 350 2 Bot 16 (Blue) 350 2 18 2 17 (Purple) 450 2  Lobster 31 (Blue) 450 2.5 31 (Green) 600 3  Gorilla 38 3 12 300 2 Giraffe 38 4 36 550 4. and design five walking automata. we enforce a symmetry relation between explore a three-legged design for which our optimization discovers a left and right legs to reduce the complexity of the design. shown in Fig. efficient. led to a slow propulsion of the center of mass. as well as 4 phase offset. 6a. the walking optimization companying video). but now in each iteration. 7a. This strategy of opti. we increased the weights selection)-to-continuous (Gaussian) sampling. CMA-GMM To validate the proposed alternative data- driven sampling strategy for our linkages. our optimization automatically discovers a gait with pk = µm + (x Am )E m (6) a lowered center of mass (Fig. While our initial design (Fig.        No. less dynamic gait that uses the body as a third leg. 6b (a). bottom. it makes sense It is worth noting that the optimization of the structural parameters that Gaussians with higher probability are sampled more. top) was also not able to walk. each comprising 6 bars and 10 pa- each initial user design. consists of two configurations. making for a total of 40 parameters. 6 Experiments the optimization discovered a new. for two of which we also create As an additional observation. a lobster automata. drawn from three the 4 phase offsets and the 8 velocity profile parameters. metry relation between the two legs. on average. To obtain a less dynamic gait. a sample is draw as: configuration for the optimization. Timing # Params CMA-GMM Time  Average CMA + GMM Name nents Mecha. but the gait is very This process is transparent to any other workings of CMA. the set of parameters exposed to the optimization consist of legs and the middle leg. not produce a walking motion for this character. i. when enforcing a sym- physical prototypes. With this procedure. Its four legs consist of 6 bars each.e. and Giraffe We designed three additional each optimization strategy for the Dog automaton.e. 6a. the initial design. Trial CMA + GMM Robot # Compo. leading to a more efficient gait. Hence. timing parameters leads to a walking motion. Gorilla. CMA-GMM converges Although conceived with the intent to walk. and the starting Σm . As suggested by Auger and Hansen [2012]. While this is not a problem in simulation. By where N is the number of parameters that are optimized in parallel. the optimization discovered a parameter set that leads to a succesful and aesthetically Dog Our first example is a quadrupedal automata with dog-like pleasing gait (see Fig. The initial design created by the user. For legs with identical structure. which the hind leg pushes while the front leg pulls.

we adopted a stochastic results for geometrically more elaborate models to better portray sampling strategy to search for a global minima. While 3D printing the robots is technically tinuation methods that incrementally increase the difficulty of the possible (ignoring the motors). However. for instance. initially support the weight of the mechanism but are progressively . we reran the optimization for the dog charac- of mass. an alternate locomotion mode is discovered. c) allowing the body to also change its shape. However. create walking automata from initial non-walking designs. b) a dynamic walking motion is found if only the leg parameters are optimized. (b) optimized automata (a) (b) (c) (b) Crab Model: a) the initial configuration is unable to walk. Physical Prototypes To validate the feasibility of our designs. we observed agreement be. As can be seen in the accompanying video. we would like to experiment with con- used for optimization.(a) (b) (a) Dog Model: (a) initial configuration that is unable to walk. although each one results in a walk with a smooth symmetric motion while balancing the heavy successful walking motion. each run Our system is able to adapt the front linkages and make the automata converges to a different solution. We used metal bolts and quicklock rings for the pin joints physics-based simulation framework is used in conjunction with as well as Dynamixel MX-64 servo motors and a Robotis CM-700 an intuitive set of objectives to measure the quality of the walking controller board for actuation. We conjecture that this behavior is owed neck (Fig. However. 7c. pin joints may break due to internal problem to increase the chance of finding global optima. Therefore. simplified geometries ments suggest that the optimization landscape exhibits a multitude that are consistent with the fabricated prototypes were eventually of local minima. Finally. is to employ external hand-of-god forces that sufficiently robust. One way strains. there contacts—is inevitably approximate. The framework takes advantage of learned valid linkage Although our simulation—and especially the treatment of frictional spaces to explore the parameters more efficiently. A ness. Laser-cutting linkages and using metal pins proved to be to do this. 1. 7c. All Our work succesfully demonstrates that computational design can body parts were laser-cut from a plywood material with 5mm thick. with a long neck. magnitude). We explore how Robustness To investigate the robustness and convergence prop- the optimization adapts to asymmetric design with a shifted center erties of our method. Figure 6: The two automata for which we fabricate results. it should be noted that we show simulation Due to the non-linear nature of the problem. we show our Giraffe-inspired automata. the initial ter with three slightly perturbed initial guesses (by 5% in relative design for the automata falls over and hits the head (Fig. As can be seen in the video. The resulting walking motion for motions as the mechanical structure of the automata is automatically these two prototypes can be observed in the accompanying video. 7 Limitations And Future Work we created physical prototypes for the Lobster and the Dog charac- ter. to the complexity of the problem and the randomized nature of our evolutionary optimization scheme. adapted. Finally. representative photographs of which are shown in Fig. are limitations to our approach and these are only the first steps. bottom). top). experi- the artistic intent of the characters. tween the walking motion of the prototypes and their simulated counterparts.

In red are inital configurations. (c) Giraffe: Heavy neck makes the optimization adapt the automata-linkages accordingly. (a) Grandpa-Bot: A human-like automata that exploits a walker for additional support. while blue are optimized walking automata. Figure 7: Various automata produced by our system. . (b) Gorilla: A three-legged design with each leg alternating between support and swing-through phase.

Flexible muscle-based locomotion for bipedal well as the static and dynamic friction coefficients of the fabricated creatures. 2011. S TEED .. material and floor — currently these are generic. mization process to more thoroughly explore the parameter space. M. 4. G EIJTENBEEK . AND G ROSS . 33. K IM .. 2009.. M. This strategy C OROS . ET AL .. G AUGE .. B ÄCHER . S.. W. Springer. B. ACM 15–22.made weaker until they eventually vanish altogether. S. C. H ORNUNG . E. C OROS . 4. 2. References P R ÉVOST.. rates. J. D.. AND S ORKINE - H ORNUNG . 1994. BERG .. ACM. 2012. J.. O. AND T IBSHIRANI .. 6. ACM as fewer trials would lead to failures. Learning. J. 59:1– B. 31.. R. To improve motion quality in general. K LEINBERGER . C OROS . In ACM Trans... Make it stand: Balancing shapes for AUGER . Graph. 31. G.. The annals of statistics 6.. R.. and so when fabricating the robots these conflicts are learning. To decrease the discrepancy between simulated and fabricated mo. In Proceedings of S CHWARZ . T HOMASZEWSKI ... B LOESCH . In ACM SIGGRAPH 2011 papers. ET AL .. tion of free-formed free-flight model airplanes. K. D. IEEE to use as a starting point for the optimization. Trans. (Proc. Learning B ISHOP... 239–245. J. E. resolved manually by stacking linkage bars along the crank rotation L EE . Graph. KOYAMA . G U . AND G UO . W HITING .. Pteromys: Interactive design and optimiza- C EYLAN . Probabilistic reasoning for assembly-based 3d modeling. L I .. and so the user cannot directly interact with the system during H OEPFLINGER . N. 30. Spin-it: Optimizing moment of inertia for SIGGRAPH ’11.. 1978. and diverse: Set evolution for inspiring 3d shape galleries.. AND S IEGWART. M. X U . 4. S. M. A. we would also like dynamic gaits for a quadrupedal robot. 75–102. 2012. 461–464. K ALOGERAKIS . Comparison of three methods for selecting values of input vari- ricating the robots. Computational design of mechanical characters. 2014. L EE .. robot moves across different surfaces. M.... C... Estimating the dimension of a model. 33. 2009. C ALIAN . Hence. ACM Trans.. Y. H ASTIE . M. Fit ACM Trans.. 64:1–64:9. Y. H ASTIE . A warm thank you to Mélina Skouras and Rocco Ghielmini for fab. (Proc. K ARPATHY. 2014. AND VAN DER S TAP - tion. ACM Trans. In ACM Trans. K. J. E. Locomotion skills for simulated Z HU .. M. E. B. Graph. F. AND C HEN .. S UMNER . The complexity of the optimization leads to longer convergence G EHRING . Z. 58:1–58:12. C OHEN -O R . to investigate methods that allow users to control the gaits at a finer G EHRING . ACM M. For novice users.. 6. quadrupeds... AND S ORKINE . H. W. S. G RINSPUN . F RIED - Finally. W. ACM Trans. The cma evolution strategy: a comparing review. Z HANG . Evolving virtual creatures. assembly. . ACM. 29.. A MATI . VAN DE PANNE . level of granularity. TAN . 57:1–57:10. 2013. ables in the analysis of output from a computer code. 2006. (Proc. 4. B ICKEL . AND H ANSEN . G. J.. ACM Trans. M. B ICKEL . J.. Data-driven biped control. T. F OR - is akin to a child learning to ride a bicycle.. 83:1–83:12. R. AND T URK . S UEDA . AND W EYRICH . YOSINSKI . 2013. N ORIS .. articulated models. N. T. linkage-based characters. D. H UTTER . Graph. B. actuators to monitor and control the gait cycle dynamically as the In Towards a new evolutionary computation. axis. M. A. T URK . G LETTE . 3d-printing of non.. mocap sequences. J. Springer. AND C ONOVER . R. Pattern Recognition and Machine bicycle stunts... M. 2013. 60:1–60:8. D. Articulated BACHER . AND S IEGWART.. K. G. R EVERET... 35:1–35:10. Y. B. V. we could add on-board sensors and additional H ANSEN . AND L EE . 2012. V.. AND C LUNE . M. strategies and covariance matrix adaptation. 2006. Graph. T HOMASZEWSKI . Graph.. M ITRA . In ACM Trans. L IPSON . 32. R.. D. 2014.. 2014. JAMES . J. G U .. H. AND P FISTER . C HAUDHURI . 32. G. (Proc. 130:1– U METANI . 2011. ACM GECCO. X U . J. those in the simulation. AND PAULY. L IU ... Evolving gaits for physical robots with the hyperneat generative encoding: The benefits of simulation. AND KOLTUN . 4. B. L. R. Technomet- rics 21. C AL Ì . SIGGRAPH) 34. M ATUSIK .. S. J ONES . 2013. Graph. 2. (Proc. G UIBAS . The elements of statistical collisions.. 2013. 59:12. AGRAWALA . B. Y. 2. A. C. R. Optimal gait and form for Asia). we could accurately estimate the force output of the motors. Springer. F RIEDMAN . these collision conflicts must be automatically resolved.. R. Graph.. Graph.. 4. C. H UTTER ... 2010. Graph. Graph. SIGGRAPH). Graph.or inter-linkage MAN . Springer. Control of the gait design and optimization process. For example.. G. Acknowledgments M C K AY. M. 4. In ACM Trans. A. W. without which this paper would not be possible. AND B ICKEL .. 2014. SIGGRAPH). T. C OROS . K. 127:1–127:10. SIGGRAPH) 31. T. M. 32.. 6. T. 2013. AND I GARASHI . 827–848... Motion-guided mechanical toy modeling.. B. K AUTZ . J. L... T IBSHIRANI . our simulation has no knowledge of intra. Designing and fabricating mechanical automata from Trans. Trans. as PEN . the legs of our fabricated lobster are broader than ACM Trans. D. Graph. L. M. spinnable objects. 31. Y. S. S. 4. Graph. S. L EFEBVRE . AND VAN DE PANNE . AND P OPOVI Ć . and could allow the opti. IEEE ICRA. 2012.. SIGGRAPH). N. 81:1–81:10.. 2011. L IU . O. C. S CHDMIT. 129:1–129:8. N. This could be investigated with additional style H OEPFLINGER .. Tutorial CMA-ES: Evolution 3d fabrication. Computational design of A. 4. or by pre-computing a library of walking templates tomatic discovery of agile gaits for quadrupedal robots. M.. S IMS . M EGARO . J. ACM Trans. B ECKMAN . and is an area of future work. vol.. 2012. (Proc. B LOESCH . S. 1979. K.. vol. Graph. Towards au- objective terms.. K. WANG .. S NYDER ... ACM Trans. TAN . Graph. Fabricating articulated characters from skinned meshes.. AND L IU . J.. C. 4. animal locomotion. W. swimming creatures. S. 130:8. ACM Trans. W HITING . C OROS . SIGGRAPH WAMPLER . In ACM SIGGRAPH. R. L. ICRA. H.