A Summary of Team MIT's Approach to the Amazon Picking Challenge 2015
The Amazon Picking Challenge (APC) , held alongside the International Conference on Robotics and Automation in May 2015 in Seattle, challenged roboticists from academia and industry to demonstrate fully automated solutions to the problem of picking objects from shelves in a warehouse fulfillment scenario. Packing density, object variability, speed, and reliability are the main complexities of the task. The picking challenge serves both as a motivation and an instrument to focus research efforts on a specific manipulation problem. In this document, we describe Team MIT’s approach to the competition, including design considerations, contributions, and performance, and we compile the lessons learned. We also describe what we think are the main remaining challenges.
From the introduction of the pallet jack in 1918 to the irruption of KIVA systems in 2003, warehouse automation has seen a wealth of advancements in the last century. The story is one of removing unstructuredness in material handling, in part by intelligent and efficient packing or palletizing. To this day, the greatest limiting factor is in the ability to handle, with speed and reliability, individual objects presented with non-trivial packing density. Of particular interest are the tasks of picking and re-stocking objects from shelves, bins, or boxes, where grasping remains an unsolved problem.
The economic value of robust and flexible object picking and re-stocking in warehouses is large and hard to quantify. Half a billion new pallets are made every year, and Amazon alone sold 426 items per second in the peak of 2013 Christmas season, all picked and boxed by hand.
The Amazon Picking Challenge (APC) , organized by Amazon Robotics, and advised by expert academics in the field of robotic manipulation, aims at focusing research efforts at addressing this very timely and potentially transformative technology of shelf-picking and self-restocking.
The 2015 challenge, described in detail in Section 2, is a first step in that direction, and tasked participating teams with the development of a fully automated system to locate and pick a small set of objects cluttered by other objects and the surrounding shelving structure. This paper describes Team MIT’s approach, who finished second among 30+ competitors.
2 The 2015 Amazon Picking Challenge
The 2015 Amazon Picking Challenge posed a simplified version of the general picking and re-stocking problem. Teams were given a space of 2x2 meters in front of a shelving unit lightly populated with objects, and 20 minutes to autonomously pick as many items as possible from a list of desired items.
Objects ranged in size, shape and material, and their arrangement was adequately simplified for a first challenge of this kind. The teams were provided with a list of 25 items prior to the competition. The exact 12 items that were to be picked from a shelf were unknown until 2 minutes before the competition and their poses and configurations were only discovered by the robot on run time.
2.1 The 25 Items
The objects used in the competition are depicted in Figure 2. According to the challenge organizers, these were selected to span a wide range of sizes, shapes and materials, and are representative of a significant portion of the item transactions that are handled on a daily basis in an Amazon warehouse. The items were chosen with the intent to provide the participants with realistic challenges such as:
Large sized objects. Items that may not fit into regular gripper spans or that may collide with the shelf during the extraction process, such as large boxes and books.
Small objects. Items whose size require high location accuracy for picking, but that are small for perception to recover accurate models.
Packaging. The reflective packaging on some of the items complicates perception, especially for depth sensors that rely on time of flight or optical triangulation.
Deformable shapes. Some items are non-rigid, which requires perception and manipulation techniques that can cope with the variability of their shapes and compliance.
2.2 The Shelf
The competition shelf is a standard Kiva pod used in Amazon warehouses, a rigid structure comprised of individual bins. To reduce the reachability requirements, the competition is constrained to the 12 bins in the center, as shown in Figure 3, which defines a cuboid of 1 meter high by 87 cm wide by 43 cm deep.
The structure has tolerances, asymmetries, and construction artifacts that deviate it from a perfect array of walls and shelves. These are the most relevant ones:
The walls and shelves are not equi-distributed. This introduces differences in the nominal size of the openings of each individual bin, with height ranging between 19 and 22cm, and width between 25 and 30 cm.
Each bin has a lip on the bottom and top edges, as shown in Figure 4, which impedes exposing an object by sliding it.
The lateral bins have a lip of the exterior edge, as shown in Figure 4, which impedes exposing an object by pulling on it.
The approach we describe here is based on planning of accurate end effector trajectories, which requires a detailed understanding of these deviations.
Finally, also worth noting, is the metallic bottom of the structure, which produced bad reflections from depth sensors and proved to be an impediment for accurate estimation of the location of the shelf by model fitting to pointcloud data.
The performance of a team was evaluated according to:
Picking a target item from a bin with one item: 10 points.
Picking a target item from a bin with two items: 15 points.
Picking a target item from a bin with three items or more: 20 points.
Damaging any item induced a 5 point penalty.
Dropping a target item from more than 30cm incurred a 3 point penalty and a non-target item incurred in 12 point penalty.
Some items had bonus points because of their increased difficulty.
It is worth noting that this first instance of the competition had many simplifications from the general picking and re-stocking problem, most of them to the object arrangement, and facilitated perception and frontal or top picking:
Items were placed close to the front of the bins, effectively reducing the necessary workspace of the robots.
Items were arranged next to each other, but not on top or behind each other.
Items were not tightly packed.
Items were not repeated inside individual bins.
The time allowed (20 minutes) was much larger than the average time that a human would take to solve the task (1 minute).
3 Design Philosophy
Our core design philosophy was driven both by the motivation to win the competition, as well as by looking towards a scalable solution that could eventually tackle the larger problem of picking and re-stocking in a real scenario. From the very beginning, we believed on the importance of aiming towards a system capable of picking a tightly packed set of varied objects, with speed and reliability, which drove most of our decisions.
For this competition, our design approach was based on: 1) The accuracy, controllability and reachability of a large and stable industrial robotic arm; 2) The flexibility and thin profile of custom made flat fingers, and 3) The versatility of a set of highly developed primitive actions.
Also critical was our effort towards an integrated system that did not need assembly or wiring in the competition. Our platform was self-contained except for a master external PC. Our system was supported by a heavy base that contained the robot, cameras, PCs for image processing, power adapter, compressor, UPS, and all the necessary cabling. We believe the reduction in the need for integration in the competition was instrumental to the reliability of the system.
The following sections describe in detail the hardware and software subsystems.
4 Hardware System Overview
The hardware infrastructure is based on an industrial ABB 1600ID robot arm coupled with a parallel-jaw gripper with custom-designed fingers, and integrated cameras. The particular robot model was chosen for its sufficient workspace, position accuracy, and high speed (we used a tool max speed of 1 m/s for the TCP during the competition). The robot has a hollow wrist and purpose-built canals that allow routing cables and airlines from the base to the gripper, which is important to maximize maneuverability and preventing cables and connectors from being pulled through the interaction with objects and environment. This section describes in detail the gripper, cameras, and other feedback sensors integrated in the system.
The design of the end-effector was driven by the desire to integrate in a compact solution three different picking modalities: grasping, suction, and scooping. As shown in Figure 5, these take place by the combined action of a parallel-jaw gripper with flat fingers, a suction cup, and a compliant spatula.
Gripper Selection. The gripper of choice was the parallel-jaw gripper WSG 50 from german manufacturer Weiss Robotics, with the following features:
Opening range of 110mm. Sufficient for most objects in at least two of their main axes. Unfortunately, the max opening of the gripper governs the nominal size of the gripper’s body, which we wanted to make sure could fit inside the bin both in vertical and horizontal directions.
Max gripping force of 70N. With sufficient friction, this gives enough gripping force to maneuver all objects without risk of dropping.
Position and Force control. Both were instrumental. Uncertainties in the exact location of the walls and bottom required opening and closing with force control to achieve a desired preload on the spatula.
Finally, we would like to highlight that we did not find significant advantages in available three fingered grippers due to their relatively large size.
Gripper Customization. The custom designed fingers shown in Figure 5 are comprised of high-strength aluminum plates, which provides stiffness, durability, and low mass.
One finger incorporates a hardened spring steel spatula, designed to deform elastically when pre-loaded against the shelf’s walls or bottom, which is key for inserting fingers under or on the side of objects flushed against the wall. Additionally, there is a suction system on the top finger for lifting items that are difficult to grasp.
The second finger has an integrated suction cup with an in-line vacuum-Venturi device. A vacuum sensor that gives feedback on strength of seal with the object and the valve to regulate the air flow are hosted in the 3D printed wrist.
Robust perception is a key component of the challenge. Distinguishing the identity of objects, and recovering an accurate pose is essential for a robust execution. The nature of the challenge poses significant challenges, including occlusions, tight view angles, and noise due to bad reflections.
We manually optimized the selection of camera and camera location to maximize the view angles for the target objects; avoid occlusions from robot arm, gripper, and shelf; avoid collisions with objects and environment; and safety of the device. For the sake of robustness, our system combined the use of:
Two Kinect2 cameras mounted to the base of the robot at about 1.7m from the ground, providing two well calibrated top-side views of the bin. Kinect2 uses time-of-flight to measure depth, which allows multiple cameras to operate simultaneously. It senses in the m range, with a resolution of 512x424. It has problems with plastic packaging due to multiple reflections.
An arm-mounted Intel RealSense camera, as seen in Figure 6 provides close-up, accurate depth point-clouds. RealSense uses structured-light for recovering depth. It senses in the m range, and has a resolution of 640x480.
Our control architecture is composed of:
A master computer (Lenovo ThinkStation, Intel Xeon E3-1241 CPU, 32 GB RAM, Nvidia Titan X GPU), that controls and supervises the entire system. The GPU in this system is used to run the perception system which is further discussed in Section 5.1.
Two auxiliary compact computers (Gigabyte Brix from Intel i7-4770R CPU, 16 GB RAM) integrated in the robot platform, capture, pre-process and filter pointclouds, to reduce the load on the master computer.
Also worth noting that, due to the problems with the max length of the USB connection of the RealSense camera, one of the compact computers had to be mounted on axis 3 of the robot. Ethernet connection seems to be preferable for its robustness and flexibility.
5 Software System Overview
At a high level, the autonomy of our system is governed by a series of primitive actions—involving both action and perception—as well as a centralized heuristic, acting as the brain of the system, that decides what primitive to execute based on the goal item to pick and feedback from cameras and sensors in the end-effector (Figure 7). The individual primitives take care of the fine details of the interaction with the environment, low level planning, and execution, also based on feedback from cameras and sensors in the end-effector.
The software architecture is based on the ROS framework , overseeing the interaction between perception, motion primitives, planning, and user interface, which we describe in detail in this section.
The role of perception is to find the identity and location of all the objects in a bin in the presence of occlusions, narrow view points and noise due to bad reflections. This is a challenging problem, and in our experience with trying different existing solutions, still a relatively unsolved one.
Our approach was driven by interaction with the software from Capsen Robotics , who gave us access to a GPU implementation of an algorithm that fits existing models to depth-only pointclouds. Our final approach was:
Filter points outside the convex hull of the shelf, and bad reflections returned as NaN. To ease the filtering, we transform the pointclouds into shelf frame, where walls are axis aligned. We then store the information in a point-based mask.
We specify physical constraints for object detection inside a given bin. Objects must lie within the bin’s walls and on the floor. To that end, the resulting object pose should intersect a up-shifted and thickened version of the bottom of the bin.
We feed the pointcloud, mask, constraints, and the ID of the objects in the bin to the Capsen Robotics software . This outputs an array of scene hypotheses associated with a score of log likelihood, each containing an array of item IDs and 6D poses.
We reject hypotheses with the center of mass outside the bin, since Capsen treats constraints as soft-constraints.
We select the hypothesis with the highest score. In case of multiple instances of the target item, we randomly pick one of them as the target object.
For the sake of robustness, we run this algorithm on depth images from the closest Kinect 5 times, and repeat another 5 times on depth images from the RealSense camera.
We would like to note that, although our final approach was based only on depth information, we believe now the combination of depth and RGB to be essential for a robust perception solution.
5.2 Motion Primitives
One of the main philosophical choices in our approach is the use of motion primitives rather than a straight motion planning approach. A motion primitive is an action defined by its goal of achieving a specific type of manipulation. The main reasons are the need for robustness and speed. By constraining the full capability of robot motion to a small set of predefined families of motions, we can focus on high-end performance for specific task, giving us a better understanding of the expected performance of the system our system while also speeding up planning since the plan (except for low level joint motions) is pre-computed during the development of the primitive. We proceed with a description of the primitives we implemented.
Percept. Responsible for moving the robot arm to place the arm-mounted camera in front of the desired bin, and getting out of the way from the fixed cameras. The primitive is also in charge of turning on and off the cameras to prevent IR light to cross-affect cameras. It is also worth noting that Kinect2 tends to blind RealSense because of the high-intensity IR strobe.
For the arm-mounted camera, we pre-select two advantageous viewpoints for each bin, so no arm motion planning is required on real time. Of special interest is the difficulty of having the camera mounted on the 5th axis instead of at the end-effector. This prevents the camera from colliding with objects and environment, but reduces is workspace. Planning arm motions is done by searching for an IK solution with a tolerance, rather than specifying a hard constraint on the 6DoF pose of the camera.
For most items, and for the sake for robustness, we use both Kinect and RealSense. The different technologies used by both cameras, give us a small gain by running both. For large, non-reflective objects, however, Kinect2 is sufficient.
Grasp. This primitive aims to end up with a vertical parallel jaw grasp of the target item in between the two flat fingers. It relies on known geometric information of the objects and the maximum opening of the gripper, which we use to define a set of advantageous grasp, a well as their expected finger opening. Based on the object shape and pose, the primitive decides: the trajectory of approach of the gripper, its pose before closing the fingers, and the expected finger opening after the grasp.
Rather than searching over the space of all possible grasping trajectories, our algorithm is based on the following principles:
The nominal strategy is to approach the object straight from the front of the shelf towards the detected pose of the object.
Search over the pitch and yaw orientation of the gripper in its approach to the object, to avoid collisions of the gripper with the shelf (including the multiple lips).
If the target item is too close to a wall, we make use of the spatula finger to add pretension on the wall so that it can slide on the side of the object. Note that by employing the compliant spatula, and exploiting the environment, object configurations that are very difficult to reach for traditional gripping systems, turn into very robust and easy to reach grasps.
Scoop. Primitive specially designed to make use of the spatula at the end of one of the fingers, as shown in Figure 5. The combination of the compliance of the spatula with the force control of the opening of the gripper, allows the robot to flush one finger against the floor of the bin with a controlled preload, and objects to be “scooped” between the fingers by pushing them against the back of the shelf.
Scoop is a powerful and very robust primitive that, again, makes use of pushing [5, 6] and the environment to reduce uncertainty  and extend dexterity [8, 9], which is specially apt for objects difficult to perceive or grasp because of being small, deformable, or flat.
Suction. Basic, but very functional primitive useful for objects with exposed flat surfaces. We limit our implementation to objects with horizontal or vertical surfaces.
For example, in the case of a horizontal-down suction, the process simply involves lowering the suction cup down onto the exposed face. Whether or not the target item presents a suctionable face depends upon its shape and current pose. The process is as:
The robot positions the suction cup above the centroid of the object and lowers the cup. The compliance of the cup and the force control of the gripper tells the robot when to stop the down motion.
We check whether the height at which the motion was blocked matches the expected height (from geometric object models).
The suction cup lifts, and the system checks the pressure sensor to determine whether suction was successful or not. We make a max of 5 attempts to suck the object with small variations in the X-Y plane.
Suction is specially advantageous for objects that are flat and wide. However, our choice of suction cup could not form a seal with bags or furry objects, which renders it un-usable for those objects. In general, picking up an object with suction is either very easy or impossible, with very few cases in between.
Topple. This is a helper motion primitive whose goal is not to pick, but to change the configuration of an object that cannot be picked by the other primitives. It is called in particular when the exposed face of an object is tall and wide. In that case, suction is not possible because the gripper would collide with the top part of the bin, the opening of the gripper might not be wide enough for grasping or scooping. In this scenario it may be possible to topple the object into a new configuration that will be grasp-able, suction-able or scoop-able.
Topping is implemented by pushing the object with a relatively rapid motion above its center of mass, so that the object rotates above its back supporting edge. The question of whether an object can be toppled can be answered beforehand , but deciding whether the primitive is working as expected in real time is still a difficult problem. Our work-around is call perception again after executing toppling.
Push-Rotate. Push-Rotate is another helper function that aims at turning an object to expose a graspable side. The maximum gripper opening limits the set of graspable objects and poses. For most of the objects in this iteration of the challenge, one or more of the dimensions of the dimensions was graspable with our current gripper. The combination of push-rotate and toppling allow us to deal with most corner-cases.
Push-Rotate chooses where and how to push the object to make the smaller dimension face the shelf front and generates the robot trajectory accordingly. After execution, as in the case of toppling, perception is called again to asses the new pose of the object.
5.3 Motion Planning
Each primitive is specified as a series of 6DOF end-effector poses and gripper openings through which the robot will need to sequence. To execute them we use the inverse kinematics planner from the Drake package developed by the Robotic Locomotion Group at MIT , which provides a detailed sequence of joint trajectories for the robot to follow. Given that the series of end-effector poses are already very descriptive of the intended motion and are designed with collision avoidance in mind, we do not take into consideration collisions with the shelf when generating the trajectories themselves.
At the task level, the autonomy of the system is driven by a heuristic, composed of a state machine and a prioritized list of primitive actions to try first for different object configurations.
The heuristic is in charge of processing the work order and sorting target objects by difficulty (based on the desired object and amount of clutter). It then goes one by one through target items from easiest to most difficult. For each target:
It scans the associated bin and decides the type of pose of the object;
Chooses and executes a strategy (set of primitives chained together) based on the specific object, its pose type, and its prioritized list of primitives.
Evaluates success, and based on the outcome, and the number of attempts, decides whether to reattempt with a different strategy or skip the item.
The choice of strategy is a crucial task. In our implementation, the specific priorities given to each primitive and object pose were fine-tuned with the help of numerous experiments to estimate success rates based on few object configurations. It would be desirable to develop a more systematic approach to optimize the heuristic used to decide what is the best primitive to use for a given bin configuration and desired item.
Our approach relied heavily on an accurate calibration of the location and geometry of the shelf, as well as the location of the cameras. The calibration begins with locating the base of the robot centered with the shelf and at an ideal distance from the shelf, optimized to maximize the dexterous workspace. This location had a tolerance of about 5cm.
We then used a semi-automated process, where the robot iterated through a series of guarded moves to accurately locate the walls, tops, bottoms, and lips of all bins in the shelf. A time consuming, and ideally avoidable procedure, but necessary for the open loop execution of many of the steps of the designed primitives. Finally we calibrated the cameras with the help of the robot. The Kinect camera point clouds are calibrated using AprilTags  attached to the robot end-effector.
The communication between different processes and machines was handled through the Robot Operating System (ROS) architecture with publish/subscribe structure when dropping messages was not critical, and services for communications that require reliability.
The interface between the overall system and the Drake planner run in MATLAB was through a standard TCP connection, and JSON format files as marshalling method.
5.7 User Interface
Finally, although the system is meant to be autonomous, the user interface is a key component specially during the developing phase.
We make use of rviz to visualize states of the system, and facilitate input into the system, such as positions and orientations with interactive markers, which is useful, for example to manually approximate the calibration of shelf and cameras, and manually input the object pose for testing purposes. The visualization is also useful for pre-screening planned arm motions which can prevent collisions due to coding errors. Our system visualizes: 1) state of robot and gripper; 2) trajectory generated by the planner; 3) pointclouds and object pose estimated from perception algorithm, which give us clues as to what components of the software architecture err when failures occur.
The overall plan for handling an order is visualized with a dynamic webpage that displays the order of items to pick, and current item.
Finally, for managing the execution and stopping all processes, we use Procman . Our system runs about ten simultaneous processes that are run in three different machines. Automating the starting and stopping of the system saves a lot of time, and avoids many errors that might happen during the starting protocol. We found Procman particularly superior to the standard roslaunch from ROS.
6 At the Competition
The competition was a motivating factor, especially to work on the integration and robustness of the overall system. As already mentioned before, we put a strong emphasis on minimizing the amount of integration that would need to be done on the day of the competition, which led us to use a heavy base that contained the entire system, except for the master computer.
Setup procedure. The robot was shipping as a whole unit, except for the outer structure that was holding the cameras in place. Once in the venue, we assembled the structure, and had to replace the compressor because it broke during shipping. After the system was up and running, we followed the calibration procedure and the system was ready to run.
During the design we also tried to minimize the time it would take to replace broken components. For example we had a second gripper which we could replace in just a few minutes. We did by making sure that all cables and tubes going to the gripper had a connector behind the wrist. It turned out to be a useful feature, given that we actually had to replace the gripper during the testing phase because a miss-calibration and accidentally bending one of the fingers.
The competition run. At the beginning of the 20-minute run, the gripper went into a reboot mode, most likely due heavy network traffic in the TCP/IP communication even before making the first attempt to pick the first item. We were penalized by 5 minutes, the gripper got back to normal state, we restarted the execution, and everything run more or less smoothly after that.
Our system picked 7 items out of 12 items, lightly damaging the packaging of one of them, in about 7 minutes before the motion supervision of the robot stopped the execution due to a torque overload in one of the joints. The system could have easily continued by pressing play on the robot controller, but the rules did not allow, since that would have constituted a human intervention. In total, we scored 88 points finishing second in the competition among the 30+ international teams from both industry and academia.
The picking challenge is designed to be a multi-year effort where solutions will need to keep pace with gradual increases in the complexity of the challenge. We enjoyed the experience and hope to participate again next year. We believe the competition was particularly enriching because:
The focus that comes from a real and well defined task. By formulating a specific problem it removes some of the academic freedom, but at the same time prevents many of the biases we tend to introduce in doing so, often in the direction of problems that we know how to solve or feel comfortable with.
The lack of constraints in the hardware or software was motivating and led to a wealth of approaches, almost as many as different participating teams. In future iterations teams might start converging in their approach based on previous experiences. We hope this document helps in understanding what we did and why we did it, as well as opinionating on what was instrumental and detrimental of our approach.
7.1 Future Work
Perception. In our experience it is very difficult to solve this problem without robust perception, and this will require a clever combination of depth and RGB data. We have observed many situation where either one or the other lack to provide enough information to positively locate an object. It is also important to notice the difficulty of plastic and reflective packaging, where both color and depth information fail often to recover useful information.
We believe an improvement to our current system might come from fusing multiple views. The accuracy of the robot, jointly with existing algorithms for pointcloud fusion, should give us more dense, multi-facetted and reliable pointclouds of the scene, which can lead to more reliable perception.
Calibration. Our system relies on an accurate calibration of the geometry of the shelf. That is not reflective of the real problem, so we would like to step away from the need for accurate calibration, by relying more on camera information and develop reactive versions of our primitives by closing the loop with either force or tactile data.
Thin manipulation. Probably the most exciting direction for the future for us is the focus on thin manipulation, or manipulation in tight spaces. We would like to be able to pick and re-stock objects in scenes with high object density and in situations of tight packing. We finish with the motivating example of picking a book from a shelf of many.
- Amazon, “Amazon Picking Challenge,” http://amazonpickingchallenge.org/, 2015, [Online; accessed 16-Jun-2015].
- B. Corporation, “This new truck gets under the load and lifts it,” Popular Science Monthly, vol. 6, no. 93, p. 54, 1918.
- M. Quigley, K. Conley, B. P. Gerkey, J. Faust, T. Foote, J. Leibs, R. Wheeler, and A. Y. Ng, “Ros: an open-source robot operating system,” 2009.
- J. Glover, “Capsen: Object pose detection software,” 2014. [Online]. Available: http://www.capsenrobotics.com
- M. T. Mason, “Mechanics and Planning of Manipulator Pushing Operations,” The Intl. Journal of Robotics Research, vol. 5, no. 3, pp. 53–71, 1986.
- M. R. Dogar and S. S. Srinivasa, “Push-Grasping with Dexterous Hands: Mechanics and a Method,” in IEEE/RSJ Intl. Conf on Intelligent Robots and Systems, 2010, pp. 2123–2130.
- C. Eppner, R. Deimel, J. Álvarez-Ruiz, M. Maertens, and O. Brock, “Exploitation of environmental constraints in human and robotic grasping,” The Intl. Journal of Robotics Research, vol. 34, no. 7, pp. 1021–1038, 2015.
- N. Chavan Dafle, A. Rodriguez, R. Paolini, B. Tang, S. S. Srinivasa, M. A. Erdmann, M. T. Mason, I. Lundberg, H. Staab, and T. A. Fuhlbrigge, “Extrinsic Dexterity: In-Hand Manipulation with External Forces,” in IEEE Intl. Conf on Robot and Autom, 2014, pp. 1578–1585.
- N. Chavan Dafle and A. Rodriguez, “Prehensile Pushing: In-hand Manipulation with Push-Primitives,” in IEEE/RSJ Intl. Conf on Intelligent Robots and Systems, 2015.
- K. M. Lynch, “Toppling manipulation,” in In IEEE/RSJ Intl. Conf on Intelligent Robots and Systems, 1999, pp. 152–159.
- R. Tedrake, “Drake: A planning, control, and analysis toolbox for nonlinear dynamical systems,” 2014. [Online]. Available: http://drake.mit.edu
- E. Olson, “Apriltag: A robust and flexible visual fiducial system,” in IEEE Intl. Conf on Robot and Autom, 2011, pp. 3400–3407.
- J. Leonard, J. How, S. Teller, M. Berger, S. Campbell, G. Fiore, L. Fletcher, E. Frazzoli, A. Huang, S. Karaman et al., “A perception-driven autonomous urban vehicle,” Journal of Field Robotics, vol. 25, no. 10, pp. 727–774, 2008.