This was Crawler, our initial prototype.
Updated: Aug 4
Unveiling the only mechanized directed energy deposition (DED) printer in the world.
PHOENIX, AZ, USA - AUGUST 1, 2021 - Rosotics, Inc. (www.rosotics.com), creator of the world’s most advanced metal additive manufacturing systems, details today the development, production, and completion of our initial prototype build, the Crawler mechanized printer. Honeycomb is an ambitious project dedicated to the establishment of a truly decentralized manufacturing system and its infrastructure. There exist no practical metal additive manufacturing systems to perform work on-location or at heightened scales. We believe in a future in which our ability to manufacture in three-dimensions, large or small, is accessible as needed and limited in no respect.
Inception architectural criteria at founding defined the need for a system to print in metal structures of large size, doing so at the same or better quality standard as required in industry using current methods. Initially, the corporate architecture plan was to utilize large industrial robotics to accomplish this, such as the KUKA KR 30 HA, outfitting said robot arms to conduct the well-established directed energy deposition process. In sourcing hardware for this system on which to develop our minimum viable product, we found this methodology to be exorbitantly expensive, having no practical alternatives. The architecture plan evolved to an attempt at minimizing the core print hardware as much as physically possible, and optimizing it for mobilization and distributed collaboration. Our inspiration for this move being systems of nature such as ants and bees, which are able to construct relatively large structures well through distribution, a new architecture plan was evolved by the name of Honeycomb, Crawler the first step.
We discovered a concept referred to as scale invariance. In physics, mathematics and statistics, scale invariance is a feature of objects or laws that do not change if scales of length, energy, or other variables, are multiplied by a common factor, In other words when applied to manufacturing, when the scale of what you are producing increases or decreases, the hardware involved in producing it remains the same. This remains an unsolved, somewhat unresearched problem that affects current manufacturing methodologies and technologies. A 3D printer therefore cannot produce objects beyond the scale of, at most, several meters - as approaching this boundary incurs extreme expense and complexity at the systems level. A scale-invariant production machine would serve as a powerful tool that allows for the additive manufacture of very large structures such as aircraft and spacecraft parts, wind turbines, or marine vessels, at incalculable cost and time savings. In order to achieve scale-invariance, the platform would need to mobilize across three dimensions (x, y, and z), that is, be airborne.
Prior to the design of this airborne platform, a ground-based prototype targeted at proving the concept, that performance metal printing could take place from an entirely mobile and autonomous platform, underwent a design cycle. This was the necessary first step before advancing to the active development of an airborne platform. While not a scale-invariant system in its entirety, it would be capable of mobilization across two out of three (x and y) spatial dimensions and demonstrate a virtually limitless functional build volume across these two dimensions, such not to exceed several feet in height off the ground. This prototype was referred to as 'Crawler', and was designed as capable of printing in aluminum and steel using the laser-wire directed energy deposition (DED) process. A wire feedstock in tandem with DED would allow for the practical operation and maintenance of the system as well as the performance required to meet AWS standards. It would be the only autonomous machine to exist that is able to conduct this process with any semblance of precision from an entirely mobile platform. Design and analysis of Crawler was conducted in Autodesk Inventor, with little to no topology optimization utilized. Crawler would be comprised of mostly off-the-shelf components, excluding that of print hardware, structure, and firmware in order to keep costs low and the time to manufacture short.
The procedural methodology of Crawler class was to begin with a small industrial precision 6-DOF robotic arm, fit it with a custom designed end affector to function as an injector, deliver to this injector feeds of aluminum wire feedstock and gaseous argon, affix radially to the injector the terminators of two separate fiber lasers, power these components using lithium polymer and sealed lead acid batteries, and affix all of these components to an all terrain spider-like frame with visual sensory and advanced computational capabilities. The result was a large mechanized robot with eight legs and a central body element. The central body element was comprised of three segments, the Canopy, the Core, and the Trunk. The body of Crawler would be entirely 3D printed of ABS Plastic, and printed in segments using several specifically sourced large format 3D printers. The body segments would be treated in an acetone vapor bath and painted a metallic finish. The process of structural printing for all of the required structural segments, joint fittings, and internal supports would take seven months across three large-format printers. Fully integrated, the large machine would weigh 42.24 kg (93 lbs), 3.7 ft in length and 2.3 ft in width.
The eight leg segments, each containing three actuable joints, utilized 24 high-torque ASME-04B high-torque servo motors, costing $1,674 and accounting for 12.72 kg of powertrain weight. The structure of each leg contained two main structural bodies - the proximal body, connecting to the Core segment, and a distal body, contacting the floor. The proximal body actuated lateral to the Core segment by means of a servo motor for positioning of the leg segment forward and backward. A triangular strut existed to interface the proximal body with the distal body, by means of a second and third servo, and was reinforced against the distal body by means of a distal pylon in tandem with reinforcement against the proximal body by means of a proximal pylon. For assembly purposes, these structures were mated and then closed with plate caps interfacing with each structure. The second servo actuated the distal body upward or downward against the proximal body, while the third servo actuated the distal body angularly. These servos affixed directly with the proximal and distal bodies, assembly aided by an opening in the side of each body removed and replaced for this purpose, known as the proximal and distal plates. These plates contained cylindrical segments each affixed to the interfacing strut, and closed with a proximal/distal plate cap. Both the proximal and distal bodies incorporated on the interior of their structures channels for wiring to be guided from the distal body towards the Core segment. The proximal body affixed to the Core segment using the first servo, reinforced by an anchor structure on the underside and a collection of anchor rods. Wiring of each servo enters the Core segment using exterior holes in the structure. Each joint was soldered and then jacketed under heat with a UL224 flame-retardant polymer heat-shrink. The wiring of each leg terminates in a female XT90 connector. In total, the leg connectome contains 96 soldered joints terminating in eight XT90 connectors for quick disconnect with power storage.
We knew in order for a decentralized and distributed manufacturing system to work, the individual agents of the system would require full, complete visual autonomy. Not only would complete autonomy be required, but this autonomy would need to come at little power draw or hardware footprint. A system capable of processing visual data, as well as to perceive a world of depth we determined to be best served by a stereo camera (a similar approach as the human visual system, employing two eyes) and a powerful COTS SoC device. For computation, we incorporated an Nvidia Jetson AGX Xavier (0.65 kg), as well as two Stereolabs ZED Mini stereo cameras. The Xavier board was housed at the front of the Core segment, vented by a hexagonal pattern above on the head structure. A ZED Mini stereo camera was mounted at the front and rear of the Core segment. Four HRLV-MaxSonar-EZO ultrasonic sensors were mounted above, in the Canopy segment, for lateral sensory. The 24 servo motors of the Legs segment were controlled by Jetson Xavier through means of a slaved Arduino Mega 2560 microcontroller, interfaced with an expansion shield to accommodate all of the servo connections necessary. Jetson Xavier conducted computational operations on the visual streams as they were recorded, to conduct simultaneous localization and mapping, among other functions, in real time. Jetson Xavier was selected as the computational workhorse due to delivering up to 32 TOPS of AI performance by means of a 512-core Volta GPU inclusive of tensor cores. Rosotics believes tensor performance is important to autonomous real-time visual processing. The platform also utilized an 8-core ARM v8.2 64-bit CPU, with 32 GB of 256-bit LPPDDR4X RAM, and 32 GB of eMMC 5.1 storage. The 512-core Nvidia Volta GPU incorporates 512 CUDA cores and 64 Tensor cores, delivering 32 TOPS of deep learning performance at INT8 and 16 TFLOPS at FP16. This platform additionally contained all of the necessary ports to interface with our auxiliary sensors. Rosotics believes vision, as opposed to LiDAR, is the natural and preferred means of sensory capability. The ZED line of stereo cameras were selected, because by utilizing a stereo-based vision stack, they are able to natively sense depth and perform other important algorithms. They record real-time video up to a resolution of 2.2K, up to 100 frames per second. Each video feed was encoded in h.264 codec. The camera units also incorporated an accelerometer, gyroscope, and inertial measurement unit (IMU). Each camera was operated at a data rate of 800 Hz, a pose update rate of 100 Hz, a pose translation drift of 1.0\%, a pose rotation drift of 0.013 deg/m without loop correction, and with 6-DoF visual-inertial stereo data collection for SLAM. Each lens of each camera was set to an aperture of ƒ/2.0. Each camera interfaced to Jetson Xavier using USB 3.0, weighing 62.9 g. The two ZED cameras provided frontal and rear visual sensory for SLAM, as well as imagery for deep learning computation. The four ultrasonic sensors mounted to the canopy on both left and right provided non-visual sensory for SLAM. A third camera of the thermal sense, an Adafruit AMG8833 IR breakout was incorporated at the rear for thermal visual data collection of the work zone.
At the front and rear of the Core segment existed manually soldered arrays of 5mm white LED flat-lensed diodes, providing lighting in low-light conditions. These LED arrays were each soldered to resistors and in turn to 5V relay modules, interfacing with the Mega 2560 microcontroller. The LED arrays were each jacketed under heat with a UL224 flame-retardant polymer heat shrink, with a total of 119 soldered joints.
A lightweight aluminum pressure vessel of welding specification was painted and incorporated directly under the canopy, housing pressurized argon at 153 bar, and plumbing transferred this Argon using a set of valves and regulators to the rear of the Core segment, where it was ejected from the exterior through tubing and delivered to the injector. This gaseous argon was continually blown over the work spot from the injector during operation at a constant flow rate to serve as localized shielding, in order to combat oxidation. The injector, affixed to the end effector of MECA500, also incorporated fiber terminators from two separate fiber lasers. These fiber lasers, two Coherent FAP800-L-16W 830NM, supplied heat for the DED process and were themselves cooled using two Arctic Freezer 11 LP-100W CPU Coolers. Laser beams were delivered to the injector using laser-grade fiber optic cables. The interior of the Core segment was further aided in airflow using an Arctic P12 Silent 120mm case fan and an Arctic P8 SIlent 80 mm case fan. The fiber lasers were powered independently using two Enersys (Hawker) Cyclon 2V/25 Ah sealed lead acid batteries (3.36 kg). Mega 2560, the LEDs, the case fans, and other auxiliary electronic components were powered using four Enegitech 9V 1.2 Ah lithion-ion batteries. Power was delivered among all of these elements appropriately using clips, connectors, buck converters, and other interfaces. To power the 24 servo motors of the Leg segment, which require great power to maintain their rated holding torque, four Turnigy 20,000 mAh 6S 12C lithium-polymer battery packs (10.52 kg) were incorporated into the Core segment for the powertrain. 1 battery was allocated to two legs, respectively. Feedstock was housed within the structure on a spool, and transferred by means of NEMA 14 0.9 deg high-precision stepper motors, driven by DRV8825 drivers. Plasti-dip rubber coating aerosol was utilized to aid in feedstock travel. A 17 gauge aluminum wire was utilized, and delivered to the injector out of the exterior using PTFE tubing. Each of these components described was supported within the structure by an aluminum underframe, which was cold-welded to the physical body and cold-welded together. The underframe further reinforces the body against axial loads and directional tension.
For maintenance and assembly, the exterior structure was removable using several sections. The Trunk Segment is a plate affixed over MECA500, mounted at an angular sense against a mounting structure of the Core Segment. The Canopy Segment is an exterior structure that fit over the length of the Core segment and the pressure set (argon vessel).
Procedurally, Crawler would first rest on the bottom of the Core segment, with legs raised. Each servo motor was positioned during integration so that this was their 'idle' state. When powered on, the legs actuated downward, lifting the body off the ground. Vision and sensory systems would perform simultaneous localization and mapping (SLAM) in order to map the environment and prepare to navigate throughout it. Crawler would identify characteristics within the environment deemed 'ideal' for printing, such as level metallic substrate surfaces and open space through which to maneuver, and then position itself based on these features for printing. For printing, the arm would actuate into position, and then execute machine code for motion, translated from a 3D model (a cartesian coordinate frame was utilized, similar to traditional 3D printers, only executed using a 6-DOF robotic arm), and each layer of the aluminum feedstock would be deposited under gaseous argon shielding flow from the injector. The robot would conclude by returning to its home position, where the legs would lower it onto its belly and thereafter lift up off the ground to idle, before powering off. During this entire process, no human is involved in the state flow.
A range of support components were acquired during this development process, such as the actual large-scale 3D printers to produce the body rapidly at low cost, custom enclosures for these printers, and their respective parts and hardware, as well as safety components and other pieces of hardware. A testament to our goal of manufacturing Stinger's airframe 100% additively, we were able to 3D print the entire body of Crawler using strong, lightweight materials. The end weight of Crawler totaled to 42.24 kg (93 lbs), 44" in length and 28" in width. The cost of development and construction, including the auxiliary costs mentioned above, totaled to $5,799.81, closing in the fall of 2020. An exclusion exists in the procurement of the MECA500 robotic arm, which was rented during the testing cycle from Mecademic.
Crawler was dimensioned of size similar to a large dog. Stinger features large rotors, increasing size through span, however the core fuselage is of similar size. The Rapid Induction Printing process introduces a new system and methodology for printing. The feature of being an airborne platform allows for true scale invariance, which is not achievable with Crawler.
We were able to cut costs significantly in production of the body segments by utilizing large-format 3D printing in advanced polymers, producing the parts in-house. This also allowed for a high degree of design modification and continual iteration. We were able to reduce costs by purchasing off-the shelf components and combining them to reach a desired effect; an example is sourcing industrial fiber laser diodes at low power and concentrating each of them to the work point, rather than spending tens of thousands for an integrated single fiber laser of similar power output. We were able to begin, conduct, and finish the entire development cycle from conception over a span of months, rather than traditionally longer time scales.
Crawler was able to demonstratively perform the DED print process at quality using small diameter aluminum feedstock, however a steel feedstock was also tested. We do not consider it to be a representative performance system however, as the Rapid Induction process iterates heavily upon the procedures and KPIs characteristic of DED. Machine learning capabilities of Crawler were fairly rudimentary in comparison with the Stinger design, and were contained to the onboard computational structure enabled by Jetson AGX Xavier. Classical simultaneous localization and mapping (SLAM), image recording, and autonomous movement formed the foundation of Crawler's firmware requirement. Crawler was limited to relatively simple geometries and could not vertically scale in size due to its grounded design, only horizontally. For this reason, Crawler was a partial step towards scale invariance, however not complete.
Impact on Airborne Design
We believe we can improve production metrics significantly for class Stinger through vertical integration. It is possible to reduce costs to defined cost criteria by reducing off-the-shelf purchases and producing components in-house, such as the PCBs. We believe we will be able to deliver a usable iteration of Stinger following a moderate time frame. In terms of performance, Stinger will provide a fundamental increase in capability when accounting for nearly every metric. The purpose built additive manufacturing process, Rapid Induction Printing, eliminates bottlenecks of Crawler's procedural frameworks, and allows for significant optimization. The addition of bionics as a foremost consideration in Stinger's design allows for new levels of technical performance to be reached. A rebuilt firmware stack leverages next-generation technologies in machine learning, wireless connectivity, and system security for application within real-world markets. An airborne platform is viewed as the final, ultimate step to achieve true scale invariance.
For more information on Rosotics, please view our published documentation on this process at www.rosotics.com/documentation.
About Rosotics, Inc.
Rosotics produces hyper-efficient metal additive manufacturing solutions that solve tougher problems in industrial engineering. A pioneer of swarm robotics, we enable today’s manufacturers to additively manufacture structures that are normally difficult or impossible to print.
Visit our Press Page