A terrain treadmill to study animal locomotion through large obstacles

In nature, terrestrial animals often move through spatially complex, three-dimensional (3-D) terrain (Dickinson et al., 2000). Small animals are particularly challenged to traverse many obstacles comparable to or even larger than themselves (Kaspari and Weiser, 1999). By contrast, many laboratory studies of terrestrial locomotion have been performed on flat surfaces (Alexander and Jayes, 1983; Blickhan and Full, 1993; Cavagna et al., 1976; Diederich et al., 2002; Ferris et al., 1998; Full and Tu, 1990; Koditschek et al., 2004; Li et al., 2012; Minetti et al., 2002; Moritz and Farley, 2003; Qian et al., 2015; Spagna et al., 2007; Spence et al., 2010), either rigid or with various surface properties (friction, slope, solid area fraction, stiffness, damping, ability to deform and flow, etc.).

Recent laboratory studies have begun to advance our understanding of animal locomotion in complex terrain with obstacles (Birn-Jeffery and Daley, 2012; Blaesing and Cruse, 2004; Collins et al., 2013; Daley and Biewener, 2006; Dürr et al., 2018; Fu and Li, 2020; Gart and Li, 2018; Gart et al., 2018; Han et al., 2021; Harley et al., 2009; Kohlsdorf and Biewener, 2006; Li et al., 2015; Olberding et al., 2012; Othayoth et al., 2020; Parker and McBrayer, 2016; Sponberg and Full, 2008; Theunissen et al., 2014; Tucker and Mcbrayer, 2012; Wang et al., 2022). Because of typical laboratory space constraints, the terrain arenas used in these studies are usually no larger than a couple of dozen body lengths in each dimension. Thus, they only allow experiments at relatively small spatiotemporal scales of ∼10 body lengths and ∼10 movement cycles. It remains a challenge to study animal locomotion in complex 3-D terrain with large obstacles at larger spatiotemporal scales.

Experiments at large spatiotemporal scales are usually realized by treadmills to keep the animal (including a human) stationary relative to the laboratory (Bélanger et al., 1996; Buchner et al., 1994; Darken et al., 1997; Full, 1987; Herreid and Full, 1984; Jayakumar et al., 2019; Kram et al., 1998; Leblond et al., 2003; Stolze et al., 1997; Watson and Ritzmann, 1997b,a; Weinstein and Full, 1999). However, only small obstacles can be directly mounted on such treadmills (Voloshina et al., 2013); larger obstacles have to be dropped onto the treadmill during locomotion (Park et al., 2015; Snijders et al., 2010; Van Hedel et al., 2002). Furthermore, such linear treadmills allow only untethered movement along one direction. Alternatively, spherical treadmills use lightweight spheres of low inertia suspended on air bearing (kugels) to allow small animals to rotate the spheres as they freely change their movement speed and direction (Bailey, 2004; Hedrick et al., 2007; Okada and Toh, 2000; Ye et al., 1995). However, the animal is tethered, and obstacles cannot be used. Omni-directional treadmills allow free locomotion in two dimensions (De Luca et al., 2009; Pyo et al., 2021), but they cannot be used with large obstacles.

Here, we created a terrain treadmill (Fig. 1A–C) to enable large spatiotemporal scale, high-resolution observation of small animal locomotion in complex 3-D terrain with large obstacles. Our terrain treadmill design was inspired by a celestial globe model. It consists of a transparent, smooth, hollow outer sphere rigidly attached to a concentric, solid inner sphere using a connecting rod (Figs S1, S2A–E). Terrain modules can be attached to the inner sphere to simulate a variety of large obstacles that small animals encounter in natural terrain (Othayoth et al., 2021). The outer sphere is placed on an actuator system consisting of three actuated omni-directional wheels (Fig. S2A–E). An overhead camera captures videos of the animal moving on top of the inner sphere (Movie 1). The animal's position is obtained in real time by tracking a marker attached to its back (Fig. S2F). This information is used by a feedback controller (Fig. 1D) to actuate the connected spheres to generate the opposite velocity to keep the animal on top (Fig. 1E,F) as it moves through the obstacle field (Fig. 2A–C, Movies 1, 2). The animal was maintained within a radius of 1 body length (4.5 cm) on top of the sphere 90% of the time (Fig. 1E,F). The animal's motion and obstacle field reconstructed (Fig. 3A,B, Movie 3) from the camera video can be used to quantify metrics such as locomotion velocities (Fig. 2G–I), antenna–obstacle contact (Fig. 3C) and antenna planar orientation relative to the body heading (Fig. 3D).

The hollow, outer sphere was composed of two smooth, acrylic hemispherical shells with a radius of 30 cm and a thickness of 0.7 cm (custom ordered from Spring City Lighting, Spring City, PA, USA; Fig. S1A). The solid, inner sphere was made of Styrofoam (Shape Innovation, Baldwin, GA, USA) and measured 20 cm in radius (Fig. S1A). The two spheres were rigidly connected concentrically using a plastic rod (McMaster Carr, Robbinsville Township, NJ, USA) with a diameter of 1.25 cm, with a near 10 cm space between the surface of the inner sphere and inner surface of the outer spherical shell. To ensure that the connecting rod passed exactly through the centers of both spheres, we made custom support structures to precisely drill through them (Fig. S1B,C). The inner sphere was secured to the connecting rod using shaft collars on both sides (Fig. S1Ai). The ends of the connecting rod had threaded holes for the outer hemispheres to be screwed onto (Fig. S1Aii). The two outer hemispheres were then mated and sealed using clear tape (3M, Maplewood, MN, USA) with minimal occlusions to the camera's view.

The actuation system design followed that of a ballbot (Fankhauser and Gwerder, 2010; Kumaga and Ochiai, 2009; Nagarajan et al., 2014), but inverted (Fig. S2A,B), and consists of three DC motors encoders (Pololu, Las Vegas, NV, USA) mounted on a laser-cut rigid acrylic base (0.6 cm thick, McMaster Carr), with a set of omni-directional wheels (10 cm outer diameter, Fig. S2C–E; Nexus Robot, China) actuated by each motor. Each set of omni-directional wheel has two parallel wheels that can rotate like a normal wheel about the motor axis. On the rim of each parallel wheel are nine rollers, each of which can rotate about an axis that is perpendicular to the motor axis and tangential to the wheel rim (Fig. S2C). We coated these rollers with a layer of protective rubber (Performix Plasti Dip) to reduce their chance of scratching the transparent outer sphere. The three motors were equally spaced around the circular base (Fig. S2D) and tilted by 45 deg (Fig. S2E). The tilt angle was chosen based on the size of the base to allow each omni-directional wheel to be perpendicular to the sphere at the point of contact (Fig. S2E), which reduces vibration and simplifies actuation kinematics. Each DC motor was powered by a 12 V, 30 A DC power supply (Amazon, USA) and had an encoder to measure its rotational speed for feedback control.

We used an overhead camera (PointGrey Flea3) to record animal motion in real time. We modified existing automated tracking methods to track the animal's body and antenna movement. We attached an ArUCo marker (Garrido-Jurado et al., 2014) on the animal body to track its pose (Fig. S2F). We chose ArUCo markers because they allow real-time tracking required for fast actuation to keep the animal on top of the treadmill and can be used to infer its position and orientation using only one camera (Fig. S2F). Prior to each experiment, we adjusted the camera position and lens focus to ensure that the topmost point of the inner sphere was in focus. We then calibrated the camera with a checkerboard pattern using ArUCo software. The ArUCo marker on the animal from the calibrated camera video was tracked in real time at 50 Hz.

To keep the animal centered on top of the inner sphere as it moves through the obstacle field, we developed a control system (Fig. 1D) with the control loop running at 50 Hz. The terrain treadmill used a computer running Robot Operating System (ROS, version: Indigo) (Quigley et al., 2009) to record and track the animal (ArUco marker), control the actuation of the concentric spheres and collect video data.

During each iteration of the control loop, the overhead camera first captured an image of the animal moving on the inner sphere (Fig. S2F). This image was used to track the ArUco marker attached to the animal to measure its position in the camera coordinate system in real time (see above). Next, the current animal position was estimated from the previous position estimate and current position measurement using a constant velocity model Kalman filter (Harvey, 1990) (Materials and Methods), which reduced measurement noise and improved estimate accuracy. To keep the animal centered, we used a PID position controller to minimize the animal position error, which is the difference between the animal's estimated position and the center of the camera viewing area (i.e. the point of the inner sphere directly below the camera's line of sight). Whenever the animal position error exceeded a threshold of 0.07 body length (3 mm), to reduce it below the threshold, the PID position controller used the position error to calculate the desired translational and rotational velocities of the outer sphere's highest point, which is defined as the control effort. To rotate the outer sphere with the desired velocity, we used actuation kinematics (Eqns 1–3) to calculate the necessary rotational velocities for the three actuator motors. Finally, the calculated motor rotational velocities were sent from the computer to an Arduino Due microcontroller, which was used to actuate the motors via H-bridge motor drivers (L298N, DROK) to generate the desired rotation. For each motor, we also implemented a PID velocity controller, which used the measured the motor speed from the motor encoder to ensure that desired motor speed was reached (Fig. S2G).

We configured the control system to be robust against loss of real-time tracking of the animal marker (see Results and Discussion, ‘Treadmill performance in keeping the animal on top’ for various observed reasons). When the animal marker was not tracked, the Kalman filter estimated the animal position by predicting its motion from previous position estimates without the measurement step (see Supplementary Materials and Methods) and allowed the control system to actuate the treadmill and center the animal, for up to 15 continuous frames (0.03 s). However, if the marker was not detected within 15 continuous frames (0.03 s), the treadmill control system would stop using the position estimate from the Kalman filter, deactivate the PID position controller, and stop the treadmill actuation until the animal marker was detected again. This prevented the animal position error (and hence the control effort) from increasing in an unbounded manner from unreliable marker tracking and position estimates and avoided drastic motion of the treadmill.

Several aspects of the treadmill must be tuned to maximize the performance of keeping the animal on top. First, an appropriate lens focal length and shutter time should be chosen to obtain images with minimal blur for reliable marker tracking. In addition, camera frame rate and resolution should be adjusted so that marker detection rate matches the control loop frequency. A higher frame rate will generate video frames too fast and not all frames will be processed by the ArUCo marker tracking. Although higher resolution images are desirable for studying finer features of animal locomotion, because they take longer to be processed by the marker tracking program, the smallest resolution that satisfies requirements such as desired control loop frequency and robust marker tracking is recommended. With the camera placed 1 m above from the top of the inner sphere, we used a 16 mm lens (Fujinon) to obtain a view of sufficient resolution (688×700 pixels) and a 5 ms shutter time.

Furthermore, the Kalman filter parameters should be tuned to ensure that the animal's position is estimated smoothly even when the animal suddenly accelerates or decelerates (see Supplementary Materials and Methods). Finally, gains of the high-level animal position PID controller and low-level actuator motor velocity PID controllers should be tuned to generate the desired response characteristics such as low overshoot and quick response time (defined as the time to reduce a large peak position error from its peak to below 3 mm; see Fig. S3C). With tuning, the actuator system can rotate the sphere to achieve desired rotation trajectories fairly accurately (Fig. S2G, Movies 1, 2).

To demonstrate the treadmill's ability to elicit sustained free locomotion of the animal while it physically interacts with the terrain, we implemented a cluttered obstacle field on the treadmill with tall pillars of a square cross-section of 1.2 cm side length, with gaps between adjacent pillars smaller than 60% of the animal’s body width (Figs 1C, 2A; Movie 1). Each rectangular pillar (∼5 g) was made of Styrofoam and covered with cardstock on two opposite longer faces. We then inserted one end of a toothpick into the pillar and glued them firmly into place. The other end was then inserted into the Styrofoam inner sphere and the pillar was firmly glued to the inner sphere using hot glue.

Following this, to develop a pipeline for measuring and reconstructing animal's physical interaction with the obstacles, we created an obstacle field with sparsely distributed cylindrical pillars (Figs 1B, 2D, 3A; Movie 3). Each circular pillar (∼7 g) consisted of a plastic tube of height 7 cm and diameter 1 cm, filled with polystyrene foam. To generate an infinitely repeatable obstacle field, we placed the pillars on the inner sphere in a soccer ball pattern (Fig. S1D). At both ends and midpoint of each edge of the soccer ball pattern, we installed a pillar normal to the spherical surface, with each pillar 4 cm apart from one another (Fig. S1D). We installed the pillars using technique described above. The two ends of the supporting rod passing through inner sphere served as two additional pillars.

We used adult male discoid cockroaches (Blaberus discoidalis Audinet-Serville 1839, body length=4.5 cm, body width=2.3 cm, mass=2.5 g, N=5 individuals for sparse obstacles) to test the treadmill's ability to elicit free locomotion, keep the animal on top, and measure animal–terrain interaction over large spatiotemporal scales. We first put the animal inside the outer sphere and then sealed it. To pick and place the animal onto the inner sphere, we attached a square magnet (16 mm side length, 3.5 g, 1.4×body mass) on the animal's dorsal side, with an ArUCo marker attached to it for tracking (Fig. 3A; Fig. S2F). The added mass from the magnet (3.5 g, 1.4×body mass) was comparable to or less than that of backpacks used in previous studies of dynamic locomotion of discoid cockroaches [e.g. 0.8×body mass (Spence et al., 2010), 1.1×body mass (Han et al., 2021) and 5.2×body mass (Jindrich and Full, 2002)]. Thus, it should not significantly affect the animal's locomotion. We used a larger magnet to pick up and move the enclosed animal from the bottom of the outer sphere to the top and dropped it onto the inner sphere. We then started the control program to keep the animal on top and camera recording for analysis during post-processing. We collected a total of 12 trials using five individuals for sparse pillar obstacles and 17 trials for cluttered pillar obstacles. Data shown throughout the paper are from trials using sparse obstacles. Occasionally, when the animal stopped moving for an extended time, it was elicited to move by perturbing the treadmill manually, which often caused the animal to move outside the camera's field of view.

To measure the animal's movement relative to the sparse pillar obstacle field during post-processing, we first measured the movement of the sparse pillars (i.e. treadmill rotation) relative to the camera. We attached 31 ArUCo markers to the inner sphere, with one each at the center of hexagonal and pentagonal regions of the soccer ball pattern of the sparse pillars projected on the sphere (Fig. S3A). We then created a map of these markers using the ArUCo marker-mapper application. Because each marker and its four corners were fixed relative to the coordinate system attached to the inner sphere (sphere coordinate system) (i.e. T₃ is known, Fig. S3A), tracking a marker on the sphere (i.e. T₁ can be measured; Fig. S3A) enabled computing the relative pose between the sphere coordinate system and the camera (Fig. S3A, T₄). When more than one marker on the sphere was detected, the relative pose of sphere and camera could be computed by solving the perspective-n-points problem (Lepetit et al., 2009), which estimated camera pose from a known set of 3-D points (marker corners) and the corresponding 2-D coordinates in the camera image. The ‘solvePnP’ program in the image processing toolbox of MATLAB was used for this purpose. Because the animal's position and orientation relative to the camera (Fig. S3A, T₂) were directly available from tracking via the calibrated camera, the animal's pose relative to the sphere coordinate system and hence relative to the obstacle field could be calculated (Fig. S3A, T₅). Because the ArUCo marker attached to the animal was not necessarily at its center of mass, a constant position and orientation offset was manually determined and added (following the procedure in Othayoth et al., 2020). Because the animal's body orientation was mostly horizontal when moving in the sparse obstacle field and 3-D reconstruction using a single overhead camera may result in large body roll and pitch estimate errors, we assumed zero body roll and pitch when reconstructing its motion.

In addition to marker-based tracking from the video during post processing, the animal's position error and total speed in the horizontal plane (magnitude of the vector sum of fore–aft and lateral velocities) can also be estimated from the Kalman filter's position estimate and motor actuation velocities both in both real time and during post-processing. However, this latter method cannot account for errors induced by potential slip of the omni-directional wheels. In addition, this method alone cannot provide the animal's position in the obstacle field or the orientation of the sphere, because the PID position controller in this method only uses the animal's position information. Hereafter, all animal position and velocity data have been estimated from marker-based tracking, unless otherwise specified. Note that because the initial prototype of the treadmill did not have markers on the inner sphere (Movie 1, left), for trials obtained using it, the animal's speed on smooth sphere was estimated by measuring the motion of features such as dents on the sphere.

We used DeepLabCut (Mathis et al., 2018) to track the animal's head and two antenna tips during post-processing. We first manually annotated them in 20 video frames each from nine trials and used this sample to train a neural network. The antenna and head positions were then tracked by the trained network (Fig. 3A). We removed obviously incorrect tracking results such as left and right antennae being flipped and obstacles detected as antennae.

Considering that the inner sphere diameter is ≈9× that of the animal body length, we approximated the spherical surface region immediately surrounding the animal to be flat and estimated the animal's equivalent 2-D planar trajectory. To visualize the equivalent 2-D trajectory, we integrated the body fore–aft and lateral translational velocities and body yaw angular velocity (Fig. 2G–I) over time, with the initial position set at origin and initial body forward axis along the x-axis. As all the trials had extended video frames in which the animal body marker was not tracked, separating the trial into multiple segments with reliable tracking, we performed this calculation and visualization for each segment separately (Fig. 2F).

To minimize occlusions and maximize reliable camera tracking, care must be taken to minimize scratching of the transparent outer sphere, and it should be wiped clean after every use to remove smudges off the surface, using a microfiber cloth with soap and water. In addition, acrylic cleaner can be used to repair small scratches and dry lubricant can be used to remove tape residue.

The terrain treadmill was capable of eliciting sustained free locomotion of discoid cockroaches through large obstacles over large spatiotemporal scales. In the longest trial, the animal moved through cluttered pillars for 25 min (≈2500 stride cycles) over 67 m (≈1500 body lengths) (Movie 1). In addition, in the sparse obstacle field (N=5 individuals, n=12 trials), the animal freely explored and visited almost the entire obstacle field (Fig. 2D,E), displaying a variety of equivalent 2-D trajectories (Fig. 2F) and a large range of fore–aft, lateral and rotational (yaw) velocities (Fig. 2G–I). On a smooth inner sphere with no obstacles, the animal moved at an average fore–aft speed of 5.6 body lengths s⁻¹ (25 cm s⁻¹) and a maximum fore–aft speed of 6.5 body lengths s⁻¹ (29 cm s⁻¹). In the sparse obstacle field, fore–aft speed was lower than 25 cm s⁻¹ during 95% of the recorded time across all trials (Fig. 2G).

During free locomotion, besides walking or running, the animal displayed a variety of behaviors during interaction with the terrain (Fig. 2A–C). For example, when moving in cluttered obstacle field, the animal often rolled its body into the narrow gap between the pillars (Fig. 2A) to traverse and occasionally climbed up the pillars (Fig. 2B). In sparse obstacle field, the animal often swept its antennae during free exploration (Figs 2C, 3D). The animal also transitioned between these behaviors and occasionally stopped moving (Movies 1, 2).

Given such a diversity of animal movements, the treadmill was capable of keeping the animal on top. For 90% of the total duration when the treadmill control system was active (50 min out of 55 min of sparse obstacle traversal, N=5 individuals, n=12 trials), the treadmill maintained the animal on the very top of the inner sphere, within a circle of radius of 1 body length (4.5 cm) centered about the image center (Fig. 1E,F). Almost all of the remaining trial duration (10%) consisted of instances when the animal was outside this circle but still atop the inner sphere without falling off (even with occasional loss of tracking). On very rare occasions (<0.001%) the Kalman filter estimated an animal position outside the inner sphere radius. This was likely because the predicted position estimate became unreliable when animal marker was not detected (see Materials and Methods, ‘Controlling treadmill motion’).

The treadmill was fast in centering the animal by rapidly reducing the animal's position error. The high-level animal position PID controller (Fig. 1D) responded to the animal's position error by reducing it to less than 0.07 body length (3 mm) within 0.6±0.6 s (Fig. S3C). Occasionally when reducing large positions errors owing to the animal suddenly accelerating or decelerating, the experimenter manually perturbing the treadmill to elicit locomotion after the animal stopped, or the animal moving out of the camera field of view and then re-appearing at the boundary, the treadmill overshot substantially after centering the animal on top (by 33±14% relative to a peak position error being minimized; N=1 individual, n=4 trials of sparse obstacles). However, we did not observe sustained oscillations of the treadmill when correcting for position overshoot. The position error increased as the animal's total speed in the horizontal plane increased (Fig. S3B). This was expected, because the faster the animal moved away from the top of the inner sphere, the farther it reached before the PID position controller actuated the treadmill to center the animal.

The treadmill was robust against failures in marker tracking. The animal's marker was not always tracked owing to temporary occlusions by obstacles, poor contrast from shadows, or small scratches, light reflection and tape seam on the outer sphere, or when the animal went out of the camera view owing to fast running or the experimenter manually perturbing the treadmill. Over the course of 55 min (12 trials) of recording of sparse obstacle traversal, there were only 16 instances when the high-level animal position PID controller stopped because the animal moved completely out of the camera view, which required the experimenter to manually adjust the treadmill and re-center the animal.

Using the reconstructed the animal–terrain interaction for the sparse pillar obstacle traversal dataset (N=5 individuals, n=12 trials), we automatically detected which pillar the animal's antennae contacted (Fig. 3A,B; Movie 3). To do so, we approximated each antenna as a line segment from its tip (Fig. 3A,B, red and blue dots) to the head marker (Fig. 3A,B, yellow dot), measured the shortest distance from points on both of the line segments to all nearby pillars, and determined whether any pillars were within 3 cm from both antennae and which among them were closest to both antennae. To assess how good this estimation was, we observed the videos to manually identify antenna–pillar contact and compared it with the estimation. The automatic antenna–pillar contact detection captured 356 out of 395 manually identified contacts (Fig. 3C, red, N=3 animals, n=3 trials) and failed to detect the remaining ∼10% (Fig. 3C, blue). It also falsely detected 111 contacts (Fig. 3C, gray). Planar orientation of both antennae relative to the body heading (Fig. 3D) suggested that the animal swept both its antennae while exploring the obstacle field.

We created an experimental platform for large spatiotemporal scale measurements of free locomotion of small animals through complex 3-D terrain with reconfigurable obstacles as large as the animal itself. Compared with existing locomotion arenas, our device increased the limits of experiment duration by ∼100× and traversable distance by ∼100. The high spatial resolution this enables will be useful for studying interaction of the animal (body, appendages, sensors) with the terrain in detail (Cowan et al., 2006; Dürr and Schilling, 2018; Okada and Toh, 2006). The large spatiotemporal scales enabled may be particularly useful for studying spatial navigation, memory (Collett et al., 2013; Varga et al., 2017) and learning through large obstacles. There may also be opportunities to advance our understanding of the neuromechanics of large obstacle traversal by combining our terrain treadmill with miniature wireless backpacks (Hammond et al., 2016) for studying muscle (Sponberg and Full, 2008) and neural control (Mongeau et al., 2015; Watson et al., 2002).

The treadmill design may be scaled up to suit larger animals such as mice, rats (Hölscher et al., 2005; Thurley and Ayaz, 2017) and lizards (John-Alder and Bennett, 1981), or palm-sized robots (Haldane et al., 2013). It may also be scaled down for smaller animals such as fruit flies (Dombeck and Reiser, 2012) or microrobots (Jayaram et al., 2020). See Supplementary Materials and Methods and Table S1 for details of design considerations for scaling up or down the treadmill. Of course, as for any treadmill or tethered experimental setup, one should be aware that the resulting locomotion will not be exactly the same as during natural free locomotion because of differences in the animal's sensory cues (Dunbar, 2004; Herbin et al., 2007) and dynamics (Buchner et al., 1994; Stolze et al., 1997; Van Ingen Schenau, 1980).

Our treadmill is only a first prototype and several improvements can be used to maximize its performance and capabilities. First, because lighting was not optimized, the pillar shadow resulted in substantial variation of brightness and contrast between the background and antennae, and because the left and right antennae are visually similar and often moved rapidly, automated antenna tracking was accurate in only ≈40% of frames after inaccurately tracked data were rejected. We will add more cameras from different views to minimize occlusions and add diffused lighting from different directions to minimize shadows. We can also increase camera frame rate to accommodate rapid antenna and body movement. These improvements will help achieve more reliable tracking of the animal body and antennae through cluttered obstacles during which 3-D body rotations are frequent.

Second, in previous kugel treadmills (Hedrick et al., 2007; Hölscher et al., 2005), the small Styrofoam sphere has a rotational inertia comparable to that of the animal's, and thus it rotates quickly in response to the animal's rapid accelerations and decelerations or large position errors without significantly affecting animal locomotion. By contrast, even though we used a lightweight Styrofoam inner sphere and obstacles, the rotational moment of inertia of all of the spherically rotating parts of our terrain treadmill is still far greater than that of the animal's (∼10⁴ times; see Table S1). This contributes to the animal's position error overshoot from the PID position controller when the animal suddenly accelerated and decelerated or when large position errors are minimized. In future, we can reduce this overshoot by including the feedforward dynamics of the rotating parts of the treadmill in the controller model, tracking the treadmill rotation in real time, and tuning the controller gains.

Third, the high-level animal position feedback control can be improved by using not only the position but also the velocity of the animal to better maintain it on top. And fourth, we can add algorithms to automatically rotate the treadmill and perturb the animal if it is found to have stopped to enable long recording without experimenter intervention.

We thank Frank Cook and Rich Middlestadt at the Johns Hopkins University Whiting School of Engineering Manufacturing Facility for assistance with mechanical fabrication, Rafael de la Tijera Obert and Hongtao Wu for installing pillars, and Noah Cowan for discussion.