Baymax is an affordable drone equipped with an Intel RealSense camera that can detect, predict, and prevent strokes in non-critically injured patients. It incorporates computer vision, batteries, and other optimizations from robotics to provide in-home health monitoring comparable to a nurse. Baymax costs under $1000 and uses a quadcopter design with the RealSense camera and an Intel NUC computer to run tests based on the NIH Stroke Scale and monitor patients' BMI and activity levels to predict and prevent strokes.

1. Intel Cornell Cup 1 REV10192015
Application Form
Baymax - An, adaptive, assistive, and predictive drone for pervasive patient care
1. Project summary
The need for medical assistive technology has become increasingly prevalent in the last decade,
fueled by the high cost and difficulty finding trusted in-home nursing services. More recently, growth in
robotic technology has been expedited by intelligent computer vision, latest battery technology, and
energy efficient parts. However, little has been done to bridge the field of robotics with medical
assistance in a consumer-grade and economical manner. In this paper, we propose a small scale and
affordable Unmanned Aerial Vehicle (referred to as a Baymax) equipped with an Intel RealSense depth
sensing camera to detect, predict and warn the user of impending risks against a non-critically injured
stroke rehabilitating patient. Baymax will incorporate the three aforementioned optimizations in the
robotics field to provide the level of care comparable to an in-home registered nurse within the
affordability of common household medical devices. In particular, this paper will discuss the process and
materials required to build a quad-rotor drone for under $1000 that is capable of predicting, detecting,
and even preventing an imminent stroke, and conclude by briefly discussing the potential use of this
drone for other medical applications.
2. Challenge Definition
The two primary issues Baymax will attempt to mitigate are the cost of home health care nurses,
and the lack of safety in family members or informal care-givers acting as a substitute for Registered
Nurses. The Centers for Medicare & Medicaid Services (CMS) estimates over 10,800 home health
agencies providing care for over 3.4 million elderly or disabled patients[1]. Unfortunately, the eligibility
for Home Health Benefit from Medicare is only limited to patients requiring part-time or non-continuous
patient care for fewer than 21 contiguous days [2]. To those who are not covered by the Medicare
program, the approximate cost of a home health aide is $21 per hour without extra charges for
additional services [3]; this leaves the option for full-time nurses unaffordable.
One of the more affordable alternatives to having an in-home nurse is the use of an Assisted Living
Facility, which provides a one-bedroom unit for approximately $3,000 per month [3]. To put this in
perspective however, the Centers for Disease Control and Prevention (CDC) estimates 66% of people
hospitalized for stroke are over the age of 65 years [4], and the American Association of Retired People
(AARP) reports the vast majority of Americans (89%) of ages 50+ want to remain in their own homes as
long as they can [5], which leaves an Assisted Living Facility inconvenient for the majority of stroke
victims.
The other issue, and the more common alternative to Medicare and Assisted Living Facilities, is the
use of family members and untrained caregivers helping their loved ones at home. Although there are
over 34 million family care-givers in the United States [5] who practice this, there is a lack of oversight,
prompt response, and experience that would otherwise be offered by a formal Registered Nurse [6].
Our solution, Baymax, is a predictive, adaptive, and assistive drone that will aid these patients or
elderly citizens who require frequent monitoring of routine health stats. In order to reduce the
dependency of stroke victims on Registered Nurses and informal caregivers, we will design a robot that
performs routine checkups, reacts immediately to sensory health data, performs analysis on historical
data; sending aggregated statistics to a doctor for further analysis and professional insight.
Although our solution is novel to home health care and medical assistance, there are a number of
non-trivial mechanical, electronic, and computational challenges which our team has outlined below.
Each of these will be later addressed in section 3, our proposed solution.

2. Intel Cornell Cup 2 REV10192015
2.1 Mechanical Challenges
The main challenge in sustained flight of aerial vehicles is the low energy density available
from common power sources. In addition, a few of the mechanical constraints pertinent to our
specific application include:
• Requiring a thrust sufficient to sustain the battery(s), frame, sensors, and electronics at an
eye-level height.
• Delivering at least four degrees of freedom in motion for full movement.
• Maintaining drone stability in the event of rotor failure (e.g. mechanical failure, crashing
into a wall).
• Preventing patient injury in the event of complete failure (drone falls).
2.2 Electronic Challenges
The primary concern when choosing the electronics for the drone is, most actuators used in
multi-rotor projects require 11-15 amps of burst current (as does the one we will select), so the
chosen battery(s) to use for drone must have a max-current rating of at least four times the
maximum current drawn by each rotor in addition to the current required by the RealSense
camera and the main board.
2.3 Computational Challenges
The most computationally intensive, yet rewarding part of our project is the use of the Intel
RealSense 3D camera to predict, detect, and prevent a stroke in a patient.
In order to empirically detect whether a patient has a stroke, we have chosen the eight most
significant categories from the NIH Stroke Scale [10] (NIHSS) and will assess the user on each
with a test.
The tests which we decided to implement are:
• Level of consciousness (LOC) - Determine the amount of movement he/she exerts.
• LOC Questions - Prompt for current month and his/her age and verify accuracy.
• LOC Commands - Request patient to grip then release fist, and measure response.
• Best Gaze - Test Horizontal eye movements for deviation.
• Visual - Request patient to follow the movement of Baymax using his/her eyes, and measure
the accuracy.
• Facial Palsy - Request patient to smile, and measure the symmetry of face.
• Motor Arm Palsy - Request patient to lift arm directly in front of them and determine if arm
can hold against gravity for 10 seconds.
• Dysarthria - Request patient to repeat a sentence and measure ability to repeat the
sentence.
To prevent and predict a stroke, the CDC recommends maintaining a healthy body mass
index (BMI) and at least two hours of moderate-intensity exercise [12]. Using the RealSense
camera and a connection to a Wi-Fi-enabled scale, the drone can determine if their BMI reaches
dangerous levels. Additionally, the fitness activity can be accumulated from other connected
devices that gather the data we need.

3. Intel Cornell Cup 3 REV10192015
3. Proposed Solution
In this section, we will design and develop a novel solution that will reduce the cost of home health
care and mitigate the need for continuous monitoring from a caregiver. Although the scope of this paper
only covers rehabilitating stroke victims, we will conclude by briefly discussing the potential uses of
Baymax in other medical applications as well. To convey our thought process, we will break the
proposed solution into its mechanical, electronic, and computational parts while remaining mindful of
the underlying healthcare cause we are solving.
3.1. Mechanics
The most common configurations for hovering rotorcraft are helicopters, tricopters,
quadcopters, and hexacopters[7]. In order to maintain 4 degrees of freedom, minimize the
number of rotors used, and preserve stability in the event of rotor failure, we chose an
actuation system in which the number of actuators equals the number of degrees of freedom
that we require. The reason for not choosing a hexacopter in this scenario was due to the
inefficiency in over-actuated systems (systems in which the number of inputs are greater than
the number of axes necessary to control)[7] and the excess weight required to hold six rotors
instead of 4. Unlike the helicopter or the tricopter, the quadcopter design allows for one rotor
to fail while maintaining stability in 3 degrees of freedom to safely land Baymax.
To further simplify the design process due to time constraints of this competition, we will
purchase a pre-built drone and install the necessary electronic components on it. The greatest
factor in choosing which drone to purchase is the amount of payload it would need to carry. The
items in Table 1 below list the necessary parts and their respective weights used in deciding
which drone we will purchase.
Note: The reason for utilizing the Intel NUC motherboard over a less power consuming
motherboard such as the Intel Galileo/Edison board is the requirements of the RealSense
camera. The Intel RealSense F200 camera we are utilizing requires at least a 4th generation Intel
Core™ processor running Windows 8.1 and a USB 3.0, which makes the Intel NUC the best fit for
fulfilling this requirement and handling the computing necessary for our application.
Given the three electronic components listed above to be mounted on the drone, we
needed to find a drone that was capable of carrying at least 1200g of payload. This and the
economical constraint narrowed our search to the Multistar Turnigy SK450 quad-rotor drone.
Table 2 below re-evaluates the total cost and weight (including the pre-assembled frame and
rotors) of the drone.
Figure 1 - Conceptual rendering of Baymax and various parts
labeled
Intel NUC / Mainboard
Rotors
Intel RealSense
3D Camera
Safety
Mesh

4. Intel Cornell Cup 4 REV10192015
Table 1 Payload of components, used in identifying a pre-built chassis to purchase
Name Unit Weight
Intel Real Sense Camera (f200) 360g
Intel NUC Motherboard (D34010WYB) 400g (including ssd, memory)
2 [3.7v 10aH] Li-Ion batteries 207g
Total 1,174g
Table 2 Payload and cost of each component reconsidered
Name Unit Weight Unit Cost
Multistar Turnigy SK450 Quadcopter 680g $180
Intel Real Sense Camera (f200) 360g $99
Intel NUC Motherboard (D34010WYB) 400g (including ssd, memory) $259
2 [3.7v/10aH] Li-Ion batteries 207g $45
Total 1,854g $628
The chosen Turnigy SK450 quadcopter is equipped with four Turnigy 2213 motors and four
10” DJI blades. Although the derivation of calculating the total thrust output by this motor is
feasible given the specifications of the blade and the rotor, we can estimate from previous
vendor data that the total thrust per rotor will be approximately 810g [13], for a total of 3240g
of thrust at 11.1V. In order to carry the total weight of 1854g, the motors would need to output
at 50% of their max current. This will influence the type of battery which is chosen in the
electronics section below.
3.2 Electronics
The most significant challenge in choosing the electronic components is determining the
necessary maximum discharge rate of the battery. Our chosen quadcopter uses Turnigy 2213
motors, which consume 13.7A when generating the necessary 810g of thrust [13]. Four of these
motors add up to a combined load current of 55 amps. To ensure safety, we decided to choose a
battery that has a maximum continuous discharge rate of 80A and a nominal capacity of
10,000mAh for optimal duration [14]. Due to the short duration of the drone when equipped
with a single battery and our initial budget of $1000, we decided to add another battery to the
drone.
Although the current consumed by the camera and NUC board are insignificant compared to
that of the four rotors, we will include this factor when calculating the flight time. The camera
will be assumed to consume the maximum current of USB 3.0 specifications, which is 900mA.
The NUC is shipped with a 3 amp power adapter, so the calculations below assume it consumes
the maximum 3 amps.

5. Intel Cornell Cup 5 REV10192015
Since flight-time is an important factor in our design process, we can estimate the
continuous flight time of the drone using the equation 1 below:
Equation 1 Calculation of the estimated flight time
3.3 Computing
Within the available 20 minutes of flight time, Baymax will be able to predict, detect, and
prevent a stroke from occurring. Each of the three parts of the solution is discussed below.
The most critical responsibility of Baymax is to detect an ongoing stroke. To do this, it will
run a series of tests based off the NIH Stroke Scale (NIHSS). If the sum of the results from these
tests is a high number, then the appropriate action will be taken to ensure patient safety. The
threshold result to consider dangerous can be modified with the guidance of a medically trained
professional.
For the duration of this competition, we will implement 8 tests that can empirically
determine the likeliness of our patient suffering from a stroke.
Level of consciousness (LOC)
Approach visible distance of the patient (3 feet) and measure the amount of motion
exhibited by him/her within the first 15 seconds of the drone approaching them. This can be
implemented using the pose detection feature of the Intel RealSense SDK, which has a
maximum range of 3 feet [15].
0 - Patient moved immediately
1 - Patient required close proximity (less than 2 feet) before any motion occurred
3 - Patient is not moving
LOC Questions
Approach visible distance of the patient (3 feet) and request for the current month and
his/her age using an on-board speaker.
0 - All questions correct
1 - One question correct
2 - both questions incorrect

6. Intel Cornell Cup 6 REV10192015
LOC Commands
Approach face-detection distance of the patient (2 feet) and request them to open and close
their eyes, then open and close their fist. The eye can be tracked using the landmark
tracking feature of the SDK, and the fist position can be identified using skeletal tracking.
Both of which are well documented in the Intel RealSense SDK Design Guide [15].
0 - Both tasks performed correctly
1 - One task not performed
2 - Both tasks not performed
Best Gaze
Approach face-detection distance of the patient (2 feet) and request the patient to move
their eyes horizontally. We will use the landmark tracking feature to determine the X and Y
coordinates of each eye as it moves horizontally. This can be used to determine the level of
nerve damage in the eye.
0 - Normal
1 - Values deviating (partial palsy but forced deviation not present)
2 - Total gaze paresis.
Visual
Approach visible distance of the patient (3 feet) and request him/her to follow the
movement of Baymax with their eyes. Baymax will slowly move in 2 large vertical circles,
during which it will measure the visual fields of the user by tracking the position of their eye.
0 - No visual loss
1 - Partial hemianopia
2. Complete hemianopia
3 - blind
Facial Palsy
Approach visible distance of the patient (3 feet) and request the patient to make a visible
smile. We can then use the landmark and expression detection of the RealSense SDK to
measure any asymmetry in the patient’s face. This can be done by comparing the Y-axis
values of the landmark points in each hemisphere of the face.
0 - Symmetrical movements
1 - Minor paralysis
2 - Partial paralysis
3 - Complete paralysis
Dysarthria
Approach visible distance of the patient (3 feet) and request the patient to repeat a simple
randomly generated grammatically correct sentence. We will use the two built-in
microphones in the RealSense camera and the Carnegie Mellon Sphinx speech recognition
API [16] to measure the comprehensibility of the sentence and compare it with values of a
coherent individual. These values can be optimized with the aid of a medically trained
professional.
0 - No aphasia
1 - Mild loss of fluency
2 - Severe loss of fluency
3 - Mute, global aphasia

7. Intel Cornell Cup 7 REV10192015
Motor Arm Palsy
Approach visible distance of the patient (3 feet) and request the patient to lift their arm
directly in front of them and hold it for 15 seconds. We will then measure if arm begins
sagging within 10 seconds using the hand tracking feature of the SDK.
0 - No drift
1 - limb holds 90 degrees but not for 10 seconds
2 - limb cannot get to 90 degrees
3 - limb falls
In addition to detecting, we can also predict an imminent stroke from happening with
routine checkups of the patient’s body mass index (BMI) and regular monitoring of physical
fitness. Although we don’t have the necessary sensors to gather this data from the drone, we
will communicate with a commercially available Wi-Fi-enabled scale, such as the Fitbit Aria [11]
and connect with a fitness monitoring device such as the Fitbit Watch [11] and pair the two
sensory data to provide intelligent feedback to our patient when a change in health habits is
necessary.
4. Performance Measures
There are a number of mechanical tests which can measure the efficiency of the robot:
Table 3 Measurement of deviation in expected values
Test Name Expected Value Actual Value % Error Description
Total Payload 1174g Measure weight of all
components
Max Payload 3240g Max weight added while
drone is able to sustain in air.
Load Current 13.7A Measure an ammeter
connected to rotors.
Flight-Time at 50% load 23 min Measure the time from full
charge until drone cannot
sustain in air anymore.
Additionally, there are a number of measurements we can take to determine the effectiveness of
our drone as a substitute for in-home patient care. Each of the tests we chose from the NIHSS can be
evaluated on a mock patient experiencing stroke symptoms and compared to the stroke scale score of a
Registered Nurse (RN) and an informal care-giver.

8. Intel Cornell Cup 8 REV10192015
Table 4 Measurement of efficiency in medical assistance
Test Name Nurse Drone Informal Care Giver % Error
Level Of Consciousness (LOC)
LOC Questions
LOC Commands
Best Gaze
Visual
Facial Palasy
Motor Arm Palasy
Dysarthia
5. Timeline & Milestones
Date (by end of month) Milestone Status
October • Predict some of the challenges
• List the necessary benefits / features
• Design measurements for testing
• Identify parts needed to build drone
• Create 3D Model prior to ordering parts
Done
November • Assemble the drone
• Mount camera
• Begin work on SDK and complete first NIHSS Test
Pending
December • Attain fully autonomous drone movement
• Complete Test 2, 3, and 4
• Begin comparing NIHSS test results with other care-
givers.
Pending
January • Complete test 4, 5, and 6
• Continue testing NIHSS test result with other care-
givers.
Pending
February • Complete test 7, and 8
• Evaluate usefulness of all 8 tests against the
cumulative NIHSS test.
• Connect with Fitbit Aria scale for Body Mass Index
(BMI) calculation.
Pending
March • Test aggressively in the UMass Lowell NERVE (The
New England Robotics Validation and
Experimentation) center for stress test.
Pending

9. Intel Cornell Cup 9 REV10192015
6. Feasibility & Resources Available
Our team consists of four undergraduate students pursuing Computer Science, Electrical
Engineering, Mechanical Engineering, and Biology. Our combined experience and innate curiosity greatly
increases the chance of fulfilling this project to completion. The solution proposed above outlines the
design process of an autonomous drone that could potentially reduce the dependency on in-home
Registered Nurses (RNs) for a budget of under $1000. Although the cost of manufacturing this using the
parts above is $630, we deliberately allocated 30% of the budget for potential replacement parts and
broken rotors. Additionally, there are a number of optimizations we can perform given further time and
resources beyond the scope of this project.
For the mechanics, the most significant optimization we would make would be fabricating our
own chassis and purchasing individual components instead of the completed Turnigy SK350, which
would significantly reduce the cost. Another factor we are aware of is the excess weight involved in
having two separate high-power batteries; with further time beyond the scope of this competition, we
can order a single custom lithium-polymer sack that would increase the flight-time of the drone and
reduce the payload and cost of the drone.
In the electronics, the heaviest portion of the components is the Intel NUC board. With further
time, we would attempt to remove this board and replace it with a transmitter that sends the USB 3.0
signals received from the Intel RealSense camera and send it over Wi-Fi using serial protocol for the
heavy computation to be done on an external device. This would ultimately reduce power consumption,
and increase flight time.
Our team plans to use the funding from the contest committee towards purchasing the
necessary components listed in figure 3 (such as the Intel NUC board). Additional remaining funding will
be invested in better battery technology and reserved for potential failure in parts.

10. Intel Cornell Cup REV10192015
7. References
[1] Cms.gov, 'Home Health Quality Initiative', 2015. [Online]. Available:
https://www.cms.gov/Medicare/Quality-Initiatives-Patient-Assessment-
Instruments/HomeHealthQualityInits/index.html?redirect=/HomeHealthQualityInits. [Accessed: 19- Oct-
2015].
[2]Medicare and Home Health Care, 1st ed. Centers for Medicare & Medicaid Services, 2015.
[3] Longtermcare.gov, 'Costs of Care - Long-Term Care Information', 2015. [Online]. Available:
http://longtermcare.gov/costs-how-to-pay/costs-of-care/. [Accessed: 19- Oct- 2015].
[4] Cdc.gov, 'Stroke Facts | cdc.gov', 2015. [Online]. Available: http://www.cdc.gov/stroke/facts.htm.
[Accessed: 19- Oct- 2015].
[5]AARP Public Policy Institute Fact Sheet, 1st ed. AARP Public Policy Institute, 2015.
[6]C. Ellenbecker, L. Samia, M. Cushman and K. Alster, 'Patient Safety and Quality in Home Health Care',
Agency for Healthcare Research and Quality (US), 2008.
[7]C. Raabe, 'Failure-Tolerant Control and Vision-Based Navigation for Hexacopters', Department of
Aeronautics and Astronautics, Tokyo, 2013.
[8]S. M, 'Facial emotion recognition using depth data', Human System Interactions (HSI), 2015 8th
International Conference, 2015.
[9] Software.intel.com, 'RealSense - Get Started with RealSense Technology | Intel® Developer Zone',
2015. [Online]. Available: https://software.intel.com/en-us/realsense/devkit. [Accessed: 19- Oct- 2015].
[10] National Institute Of Neurological Disorders and Stroke, 'NIH Stroke Scale', 2003. [Online]. Available:
http://www.ninds.nih.gov/doctors/NIH_Stroke_Scale.pdf. [Accessed: 19- Oct- 2015].
[11] Fitbit.com, 'Fitbit Aria™ Wi-Fi Smart Scale', 2015. [Online]. Available: https://www.fitbit.com/aria.
[Accessed: 19- Oct- 2015].
[12]D. Prevention and C. Prevention, 'Preventing Stroke: Healthy Living Habits | cdc.gov', Cdc.gov, 2015.
[Online]. Available: http://www.cdc.gov/stroke/healthy_living.htm. [Accessed: 19- Oct- 2015].
[13] HobbyKing Store, 'Turnigy 2213 20turn 1050kv 19A Outrunner', 2015. [Online]. Available:
http://www.hobbyking.com/hobbyking/store/__5688__Turnigy_2213_20turn_1050kv_19A_Outrunner.
html. [Accessed: 19- Oct- 2015].
[14] Lithium Polymer Battery Specification, 'Lithium Polymer Battery Specification', 2015. [Online].
Available: http://www.batteryspace.com/prod-specs/5072.pdf. [Accessed: 19- Oct- 2015].
[15] Intel RealSense SDK Design Guidelines, 'Intel RealSense SDK Design Guidelines', 2015. [Online].
Available:
https://software.intel.com/sites/default/files/managed/27/50/Intel%20RealSense%20SDK%20Design%2
0Guidelines%20F200%20v2.pdf. [Accessed: 19- Oct- 2015].
[16] Speech.cs.cmu.edu, 'Speech at CMU', 2015. [Online]. Available: http://www.speech.cs.cmu.edu/.
[Accessed: 19- Oct- 2015].

11. Intel Cornell Cup REV10192015