Our team used weighted objectives tables and design reviews to begin our engineering process, and we prototyped many potential solutions before selecting our design. After identifying that a low mass, agility, and parallelization were ideal for this competition, we positioned high-mass components as low as possible in our compact drivetrain and heavily pocketed the aluminum plates. In order to speed up cycle time, we opted for two separate extensions for intake and output, each with several degrees of freedom to reduce drivetrain movement and allow agile mechanisms to perform fine adjustments and operate in parallel for both types of cycles. To avoid speed reductions on our vertical extension, I developed a new constant force spring mechanism to artificially increase the downward force of our extension during the hang period while maintaining high speed under driver control without using higher torque motors. To automate as much of operation as possible, we used sensors to trigger every step of our transfer, driver assists, and PID control with feedforward gravity adjustments.

Due to a six servo limit, I used a critically constrained setup with six cables arranged in a setup with diverse tension directions across 3D space. This evolved into a triangular antiprism on the outside connected to a hexagonal prism of attachment points the inside, allowing each cable to have a meaningful and necessary tension vector to enable all six degrees of freedom. As the work area was very large, I laser cut a hexagonal base to mount each component with countersunk screws from below, and used the same configurable 3D printed cable redirection block on each vertex with 3D printed V bearings. Half of these blocks had a vertical aluminum extrusion for the upper cables, and the other half released the cables from near the base. To maximize range of motion using servos with limited ranges, I implemented a unique single arm tension system where a fixed endpoint on the cable was lengthened by a V bearing on the arm as it rotated, doubling the length differential over mounting the endpoint on the arm. I used matrix operations and interpolated polynomials from experimentation to calculate angles for each servo given a target position vector and Euler rotation vector. To ensure control during motion, I added tension biases to each servo, but found that more complex tensioning systems incorporating the arm angle were unnecessary.

I wanted to leave the entire robot above and separable from the ball, so the three wheels were positioned at 45 deg angles of elevation and sprung inward to increase friction and grip. To spring them inward, each wheel module was supported on bearings and were hardstopped inwards by each other and outwards by the cylindrical enclosure. Inside each module was a 7:5 ratio of double helical gears, integrated into the omni wheels and servo attachment for a lower part count. Low profile flanged bearings were used on both sides of each rotating member to make them more robust and durable. The control system, battery, and wiring, were all integrated cleanly into the cylindrical enclosure, with paths designed into the part for the main wires, and the top is removable and allows easy access to replace the battery.

Our team used weighted objectives tables and design reviews to begin our engineering process, and we prototyped many potential solutions before settling on our design. We prioritized modularity and consistency, and achieved this by using standardized components and easily removable subassemblies. Our 3D printed parts were tested in many 3D printed materials for our parts to lighten and stiffen them over time, including nylon, PETG, and alloy 910, and we tested several types of string for our rigged extension.We opted for a virtual four bar arm to avoid a transfer mechanism and maintain the cones' orientations as they were processed. This also allowed a passthrough cycle structure, eliminating the need to rotate the drivetrain each time. Our main vertical extension used constant force springing to counteract gravitational force and the virtual four bar arm was sprung to perfectly oppose gravity throughout its range of motion, so we could focus on optimizing these mechanisms for speed and consistency.

Rather than pivoting the robot around the scanned object, I opted to rotate the object on a large turret to reduce moving mass and complexity. This involved using a 10:1 gear reduction and a laser cut gear stack to avoid 3D printing large parts. I used a rack and pinion with this specific gear ratio to map the 270 deg range of the servo to the 250 mm vertical extension. Both of these decisions balanced precision with speed, and allowed for precise encoder readings to correspond to distance measurements. On the software side, I made a simple program to remove cones from a cylindrical 'stock' piece, with each cone's vertex at the position of the sensor at that point and the base representing the distance detected. The angle of each cone stayed constant, as the field of view of the sensor, and the resulting form was a rough reconstruction of the original object after iterating through each detection.

Since the Nerf gun had very high inertia, I used a turret with a large gear reduction printed in four modular segments with dovetail joints to allow amplify the torque of a smaller motor. Again, to allow a smaller motor to control pitch rotation, I used a string and spool mechanism attached to the lighter side of the pivot position and allowed gravity to tension the string. The battery, breadboard, and wires were mounted to the turret as well, and buttons were added to turn the tracking and power on and off. The motors I had were inexpensive DC gear motors without encoders, so the tracking had to be done in relation to the current position of the gun, meaning the camera would be clamped onto the front of the gun and provide relative positioning information. Motion blur and latency were also concerns, so tracking was slower than desired, and the gear reduction further slowed yaw movement.

Stepper motors were ideal for this device due to their high precision and easy integration. Since I only had two, I used one on each axis and mechanically linked both sides of the X axis together to ensure synchronized movement and prevent a moment from developing on the side opposite the motor from asymmetric loading. GT2 belts were selected for their low backlash and high positional repeatability. Sanding the steel rods that formed the structure allowed for a tight transition fit of the bearings and reduced friction between sliding components. Each contact with these steel rods had a 3D printed compliant clamp that could be tightened and undone easily, as did the GT2 belt clamps, such that the length and depth of the grantry as well as belt tension could be easily adjusted. On the end effector, I used another 3D printed compliant clamp to hold the pen in place while allowing for easy replacement and tuning, and a small rack and pinion system to raise and lower the pen. Autonomous tracing was achieved using an edge detection algorithm that converted color images into minimal path sequences, enabling the pen to follow contours without predefined toolpaths.

For precise motor control, I used two microstepping stepper motors to drive each dial. The housing was assembled in four parts for modularity and ease of assembly, while still not exposing any internal components and leaving access for necessary external wiring. I designed it to be bounded by the smallest footprint possible between the two dials while containing the microcontroller and stepper motor driver boards and not obscuring the screen of the toy. To add additional transparency and diagnostic information, I added a small captive blue LED in the top surface whose frequency indicates the proportion of the drawing that has been completed, as well as whether the device is recieving power. The dial grippers were simply PETG printed clamps that used a small screw and captive nut to constrain the dial against a cylindrical pad opposite the screw. To generate the paths, I first split the image into contours using adaptive thresholding (to use similar relative brightness comparisons throughout the image rather than an absolute scale) and contour detection. After splitting contours at sharp angles, a node graph was created representing allowable connected contour pieces that were sufficiently close, and a heuristic simulated annealing approach was used to find a path through the graph to maximize total arc length. Unfortunately, due to slop inside the Etch-a-Sketch that I had trouble measuring, error accumulated over the drawing, but otherwise the design was successful.

Since I wanted to make the robot resistant to backdriving and retain its state without power, I used worm drives to actuate the turret and both segments of the arm. These three degrees of freedom provide the main 3D positioning capabilities of the end effector, and were made rigid using precisely toleranced bearing stacks, numerous standoffs supporting both arm stages, and a well-tensioned HTD belt. The main body was designed in three sections to make replacement and maintenance easier, as well as allowing a more enclosed and compact assembly with high mass components in a cluster near the lowest pivot position. The robot was fastned securely to the turret with four bearing stacks with flanges on both sides and a grooved gear which fixed in space to support the weight of the arm. The claw is a tightly packaged removable assembly with three servos, housing a directly driven differential gear for roll and pitch as well as the screw that opens and closes the claw. Lastly, I derived the inverse kinematic equations to control the arm angles and wrist state given a target position and wrist pitch.

To maximize the potential speed of the robot, I used one gripper module for each face of the cube, allowing for independent movement and faster overall operation. Each of these modules included linear and rotational actuation, with a mechanism designed to keep a linearly static contact point on the center of the cube's face regardless of the module's state. Given this constraint, I designed a mechanism that used a rack and pinion system where the rack was cylindrical with radial teeth meshing with a toroidal pinion, allowing the gripper to spin about its central axis without losing engagement with the pinion and vice versa. The central contact point had a hexagonal profile near its base to constrain the angular position of the gripper without locking it in place linearly or moving the support. The grippers were made of TPU with geometry such that they could slightly deform around the cube with chamfers to improve forgiveness. Six such modules, along with a hinged upper piece, formed the whole robot and allowed the cube to be removed and inserted with no disassembly. To scan the cube, two vertex-on pictures are taken of the cube, seven points of interest defining the cube's shape are selected, and the 75 visible tiles are categorized into colors using a lighting invariant method. I decided to use an existing website (Grubiks) to find the solution, and wrote the scheduling algorithm to maximize simultaneous moves and prevent the grippers from crashing during operation. Unfortunately, a circumstance outside my control during development prevented me from finishing the project and documenting a complete solve, but it is clear that the robot is fully capable of solving the cube as seen in the attached videos.

Our team used weighted objectives tables and design reviews to begin our engineering process, and we prototyped several potential mechanisms for intake and scoring before settling on our intended design. We prioritized compactness and parallelization, and achieved this by building a low profile mecanum drivetrain with pocketed aluminum plates and a passthrough transfer. Since the pieces could be handled in pairs, we used a heavily iterated dual-stage intake arm with rubber star shaped wheels to quickly grab them form the ground side by side or off of a stack using a TPU 3D printed ramp with zero ground clearance. Over the season, we developed a conveyor belt system to facilitate moving the game pieces to our transfer zone, experimenting with high-friction rubber materials to maximize grip. We ended up with a tightly constrained conveyor system below and above the transfer area that enforced game elements being directed to the right places side by side and easily accessible for the output system. To score the pieces, we used two independent plungers with high friction material in a symmetric assembly to release them quickly near the edges of the board and in any order. We used two distance sensors and solved the trapezoid formed by their lines of sight to align autonomously to the board, and our angled extension coplanar to the board enabled height adjustments without needing drivetrain adjustments.

To maintain stability when the robot encounters a vertically attached cable, I designed four sprung rollers that would deflect the cable backward one at a time while the other three supported the robot on the main cable. This seemed like the most simple and straightforward solution without an active gripping and releasing mechanism, which seemed too cumbersome and unstable. Ultimately, the four roller solution only was successful at very low speeds and failed when the robot swung along the cable too much while moving forward. I iterated on these rollers' profiles and spring forces such that they would stay in contact securely and snap back to their neutral position quickly without plastically deforming the springs, and I arrived at a decent solution but not one that performed sufficiently well. To move the robot along the cable, I used two 35A durometer compliant wheels mechanically linked and driven directly by a high speed motor to compress the cable between the two pairs of rollers. To use one servo for both the opening and closing functionality, I used a partially toothed gear and arm to both open the door at one side of the servo's range, open the floor at the other side of the range, and keep both tightly closed in the remaining middle region in the range to simplify and lighten the overall mechanism.

This was a fairly simple and straightforward construction, it consisted of four vertical motors with bevel gears driving each wheel independently. This architecture was chosen because it was easy to build and work with and didn't take up very much space laterally. I attached the linear slides using plasma cut aluminum plates symmetrically and used a coiling telephone cord to manage wiring between halves and a simple 3D printed platform for the control system and battery. To explore as many modes of control as possible, I decided to allow the driver to cycle through three control modes, each making certain motion types easier to perform. To make the wiring more secure, I prototyped a laser cut scissor linkage, but this was removed due to buckling concerns.

The launcher pivoted on one servo, which was fixed, and the spring preload was controlled using a second servo and a single arm link to change the spring's initial extension over twice the radius of the arm, defining the slowest and fastest launches. I opted for a continuous motion to launch the projectiles to facilitate back-to-back launches and simplify the control of the turret, and achieved this using a partially thoothed sector gear that would interface with a gear rack over 70% of its rotation, pulling back the plunger and simultaneously allowing the block to fall into position, and release during the other 30% to allow the plunger to snap forward, both launching the block and retaining the stack safely with the body of the plunger. I used low-friction PETG laser cut with L brackets to form the stack so the blocks could slide down easily, and a rack and pinion to simplify linera motion.

I decided to apply torques in the pitch direction using the wheel(s) and in the roll direction by moving the breadboard, the heaviest component, side to side. This was done using a rack and pinion controlled from the bottom and was able to shift over half of the robot's weight laterally by several inches. To control the robot, I used simple sensor fusion techniques with the accelerometer and integrated gyroscope data and a PD control loop, as this performed the best while experimenting. However, I found that most of the time the motors weren't able to react quickly enough to stabilize the falling robot.