This article describes features of FIRST, an approach to introduce science, technology, engineering and math (STEM) concepts to students. One unique aspect of the FIRST experience is that the reward and recognition for achievements are not necessarily gained on the field of play—excellence in design, demonstrated team spirit, Gracious Professionalism, community outreach, and more are recognized with awards. The focus of organized FIRST programs is on pre-college students {LE Comment: Please check whether the introduction of a preposition is required.}. At the college level, there are no formal FIRST programs, except for FIRST-related scholarships. However, as shown in the longitudinal study, the impact of FIRST carries into postsecondary education. There are diverse types of FIRST-related organizations and models at the college level. The article also highlights that during the semester, a sequence of team building, mentoring, project management, fundraising, and FIRST robot design topics are given by faculty, senior FIRST mentors, and professional engineers.

## Article

Bipedal humanoid robots, such as Atlas by Boston Dynamics and Valkyrie by NASA's Johnson Space Center (Figure 1), have great potential to perform a variety of complex human-like mobility and manipulation tasks, for example, climbing, crawling, traversing over rough terrain and reaching behind obstacles, when compared to their wheeled or tracked counterparts. This makes them applicable in operations in cluttered and dynamic environments that typically require making contacts with the surroundings, such as manipulation inside science glove boxes commonly found in nuclear energy facilities; in spaces specifically designed for humans, such as habitats to support future manned space missions; in operating human tools such as hand drills; and utilizing human interfaces such as switches, plugs, and door handles. The trade-off, however, is the complexity that comes with the high degrees of freedom. For example, the well-studied 2D navigation problem for a mobile robot becomes a multi-faceted research and development effort incorporating path planning, step-planning, balancing and locomotion control. How might we design and develop humanoid robots with human-like locomotion and manipulation capabilities?

In order to accelerate the research and development of humanoid robots, the robotics community has recently participated in two challenges, namely the NASA Space Robotics Challenge (SRC) in 2017 preceded by the DARPA Robotics Challenge (DRC) held from 2012-2015. With significant investments in funding the teams and providing prize money, these challenges mobilized hundreds of researchers, technologists and citizen scientists to develop solutions towards realizing practical humanoid robots in real-world environments. This article is aimed at providing an overview of both the DRC from a participant point of view and the SRC from a host team perspective. We will share our experiences from both of these challenges and point out future research and development directions.

## The Darpa Robotics Challenge, 2012-2015

On April 24, 2015, about 20 U.S. roboticists boarded the train from Tokyo’s Ueno train station for a day-long visit to TEPCO’s Fukushima Daiichi Nuclear Power Plant to observe the cleanup operations at the disabled facility as a result of the nuclear accident caused by the tsunami following the Töhoku earthquake on March 11, 2011. The trip, co-organized by DARPA (Defense Advanced Research Projects Agency) and METI (Ministry of Economy, Trade and Industry), provided the DRC finalists an opportunity to better understand the challenges in deploying practical robots to respond to man-made or natural disasters. The observations included many human workers in full personal protective equipment working all around the site, full body radiation scanners at frequent locations and operators controlling teleoperated robots being used for the clean-up.

The DRC tasks were motivated by the pressing need to respond to calamities that follow natural and manmade disasters such as the hydrogen explosions in the Fukushima Daiichi Nuclear Power Plant that happened within the first 24 hours after a tsunami triggered by the Great East Japan earthquake in 2011. From 2012-2015, hundreds of robotics researchers, practitioners and citizen scientists worked towards developing capabilities for human-robot teams to complete tasks relevant to disaster response missions such as traversing doors, operating a drill to cut a hole on a wall, flipping electric switches, turning valves, traversing uneven terrain, climbing stairs, and driving and egressing a vehicle.

The DRC Finals were held on June 5-6, 2015 with the participation of 23 qualified teams from around the world to demonstrate performance of humanoid robot-human operator teams in a disaster mission scenario.

In a simulated environment, robots performed a variety of manipulation and mobility tasks under human supervision with a 1-hour mission completion time. For greater realism, the communications between the operator(s) and robot were degraded during parts of the mission.

Through our involvement in the DRC Finals, we have learned valuable lessons on the theory and practice of humanoid robots for disaster response. We report a summary of our findings [1] : (i) It is a fact that robots experience downtimes caused by various reasons including upgrades, failures caused by degradation of hardware and failures caused by operator errors. To provide practical robot systems for realistic operations at disaster sites, it is essential to have reliable robot hardware. This includes powerful and robust universal grippers that can perform manipulation tasks. One way to mitigate hardware failures in the real world would be to design with redundancy both in terms of hardware and in terms of behaviors. Furthermore, exploring the design space to develop hybrid systems will help achieve practical humanoids. For example, the DRC first-place winner KAIST developed a robot that could perform both bipedal walking and kneel and drive based on the task. (ii) Even though we strive for best practices in software development, the number of developers with varied backgrounds working in a large team setting makes it difficult to achieve consistency and reliability in software systems. A robot operating in challenging environments needs to be programmed to have persistent behaviors to complete practical tasks. It is also critical to equip robots with safety software. Testing early and often is also important in reliable software development. On day 1 of the DRC Finals, Team WPI-CMU’s Atlas robot dropped the drill after successfully picking it up and turning it on when the robot gripper was sent an open signal by the first occurrence of a software bug.

(iii) Factoring tasks to maximize the utility of the human-robot team is essential. It is critical to pursue more autonomy in the development of robots for routine and emergency operations. The real scenes will be much more unstructured and chaotic than the simulated environments we currently use in robot testing and validation. However, current state-of-the-art only allows for short-term and partial autonomy. Models of computation can incorporate methods to improve the percentage and speed (time) of successful task completion by providing capabilities to the human operator to intervene with robot decisions at key steps. Most DARPA-funded DRC teams automated many subtasks and tasks during the development period between the trials in 2013 and finals. (iv) A rigorous validation strategy and operator training are essential. Executing a rigorous validation plan improved performance of the teams. Construction of realistic testbeds for validation and operator training will contribute new know-how to identify opportunities to speed up the task completion. Unifying multiple tasks and developing recovery behaviors will enhance the reliability of a robot by an order of magnitude through a rigorous validation strategy. Top DRC Finals teams spent the final months of their development timeline outdoors in their testbeds to run their methods repeatedly and train their operators. (v) Operators prefer to control the robot at many levels, and they want to rapidly and conveniently switch between control modalities to: command joint velocities or changes in positions to recover from failure events such as the drill tool getting stuck on the wall, command Cartesian velocities or changes in positions to nudge the robot actions to provide adjustments such as to align the robot hand with the door handle for better grasping, designate end effector targets such as desired grasps or footsteps, provide task parameters such as speed, and select tasks or subtasks to perform. (vi) Human-robot interaction with humanoid robots is a challenging research and development problem due to the complexity of the system. Supervised autonomy at the DRC (Figure 2) was only possible by operators who had extensively practiced for months, and even then errors were made. For example, operator errors were the main cause of falls by several Atlas robots during the Finals. For many, the DRC pushed the limits of humanoid robotics research efforts to new fronts from motion planning and control to perception to human-robot interaction. The DRC Finals demonstrated that state-of-the-art humanoid robots are slower than humans by an order of magnitude in performing tasks such as turning valves, using hand drills and flipping electric switches. Furthermore, the DRC robots relied heavily on pre-scripted motions and hence lacked true autonomy, a critical capability especially in performing practical tasks with complex humanoid robots, to carry out simulated tasks relevant to disaster response. We also learned that reliability in completing tasks is prohibitively low to make robots practical even for an overly simplified set of tasks. As a result, the DRC outcomes once again emphasized the need to develop new control paradigms for humanoid robots with tightly coupled perception and action loops for reliable and sufficiently rapid robot capabilities to perform tasks in dangerous, distant and daring environments.

## The Nasa Space Robotics Challenge, 2016-2017

The DRC Finals concluded by a post-event workshop during which a new challenge and call for proposals was announced by NASA. The program was designed to award two Valkyrie humanoid robots designed and developed by NASA’s Johnson Space Center, originally for the DARPA Robotics Challenge, to be hosted by two DRC participant teams in conjunction with the announcement of the Space Robotics Challenge (SRC). The SRC was a NASA Centennial Challenge requiring teams to integrate mobility, manipulation, and perception to accomplish several space exploration tasks such as alignment of a communications dish, repair of a solar array, and finding/repairing an air leak in a habitat. The SRC was motivated by the need to better understand the challenges on the path towards enabling the next generation robotic space missions using humanoid robots by leveraging the momentum in the research and development community triggered by the DRC. It is also envisioned that humanoid robots could be part of the pre-deployment missions to Mars and maintain and repair equipment and supplies to be sent in advance of the manned missions.

As one of the two host teams, the SRC provided our team not only the Valkyrie hardware platform to validate the SRC tasks but also the opportunity to expand our work on motion planning control framework from the DRC to incorporate new capabilities. In this framework, we can calculate a sequence of collision-free robot configurations efficiently by solving a trajectory optimization problem which is constrained by the kinematics requirements of the task and the collision avoidance conditions [2].

The objective of the motion planning problem, which contains a sequence of T joint configurations representing motion trajectory of a K degree-of-freedom humanoid robot system as decision variables q1:T where qt ∊ ℝK describes the robot joint configuration at the t-th time step, has the following form:

$f(q1:T)=∑t=1T((qt+1−qt)TQ1(qt+1−qt)+(qt−qnom)TQ2(qt−qnom)+Δd(qt)TQ3Δd(qt)$

where Q1, Q2, Q3 > O are weight matrices, qnom, represents a nominal posture, and Δd(qt) is the Cartesian deviation between a link’s pose at the robot state qt and its desired posture. These quadratic cost terms represent penalizations of the weighted sum on the joint displacements between the waypoints, joint configuration deviation from a nominal posture and Cartesian displacement from a link frame to a desired target frame. The first term can limit the movement of the robot and smooth the trajectory. The second term is used to push the joints to the nominal configuration when all the constraints have been met. Similarly, the third term is used to push links to specific positions and orientations.

Many constraints can be specified on the robot’s motion, which range from simple joint limits, to position/orientation constraints of the robot’s links, to collision avoidance constraints, to constraints on keeping the horizontal projection of center of mass on the support polygon. The Cartesian posture constraint plays an important role for generating a motion solution. The hardware implementation on Valkyrie requires a variety of costs and constraints to be set to generate a feasible motion solution. There are a number of general costs and constraints, such as joint displacement costs using a normal standing pose as the nominal pose, pelvis height and orientation costs, torso orientation costs, joint limit constraints, end-effector target constraints, collision avoidance constraints, and center-of-mass constraints. Figure 3 depicts a sequence of tasks while performing a simple box pick task with Valkyrie.

The SRC finals took place in June 2017 and the details of the approaches that 20 different teams used to solve the challenge tasks virtually using the Gazebo simulation environment have been revealed. The two host teams provided access to the robot, and assisted the virtual challenge winners in implementing and testing their code on the actual hardware robot.

Figure 4 presents four different approaches the SRC teams adopted to solve the first task virtually. The first SRC task involves aligning a communications dish and it requires both dexterous manipulation and constrained motion planning. Motivated by the approaches SRC teams used, our team implemented and successfully completed this task on the actual robot hardware fully autonomously using a power grasp on the knob attached to the wheel (Figure 4). To generate the motion to turn the wheel, we utilized our motion template library that is capable of storing constraints to be used by our whole body trajectory optimization motion planner. We solve the optimization problem by dissembling it into two subproblems and then solving them hierarchically. The first problem is a simplified optimization problem consisting of kinematic constraints and collision constraints. The second problem is to interpolate the kinematically-feasible path by taking dynamics constraints and smoothness into account. With this hierarchical motion planning approach, a feasible and collision-free trajectory can be generated in a reasonable time, on the order of seconds. As a host team, we provided the SRC first-prize winner, Coordinated Robotics from Newbury Park, CA, with access to Valkyrie and successfully implemented the team’s approach to task 1 on the actual hardware. This effort also demonstrated the capabilities of the current state of robot simulation environments (namely Gazebo from Open Robotics) to bridge the gap from simulation to hardware implementation.

We envision that humanoid robots will find numerous practical applications as pre-deployed space exploration assets, personal assistants in the home, nursing assistants in hospitals, and first-responders in disaster relief. The DRC and NASA SRC provided the research community with a charge to take bold steps towards realizing practical humanoid robots. The next challenge, yet to be announced, will bring this vision one step closer to reality.

Aboutthe Author

Taskin Padir is an Associate Professor in the Electrical and Computer Engineering Department at Northeastern University.

He received his PhD and MS degrees in electrical and computer engineering from Purdue University. He holds a BS in electrical and electronics engineering from the Middle East Technical University in Turkey. He is the Director of Robotics and Intelligent Vehicles Research Laboratory (RIVeR Lab). He is also the co-founder of the Robotics Collaborative at Northeastern. His research interests include supervised autonomy for humanoid robots, shared autonomy for intelligent vehicles, and human-in-the-loop control systems with applications in exploration, disaster response, personalized in-home care, and nuclear decommissioning. He led project teams for the NASA Sample Return Robot Centennial Challenge, SmartAmerica Challenge and the DARPA Robotics Challenge. Padir currently leads one of two research groups selected by NASA to develop autonomy for humanoid robot Valkyrie.

Acknowledgments

This material is based upon work supported by the National Aeronautics and Space Administration under Grant No. NNX16AC48A issued through the Science and Technology Mission Directorate, by the National Science Foundation under Award No. 1451427 and by the Defense Advanced Research Project Agency, DARPA Robotics Challenge Program under Contract No. HR0011-14-C-0011.

## Reference

Reference
1.
DeDonato
,
M.
,
Polido
,
F.
,
Knoedler
,
K.
,
Babu
,
B.P.W.
,
Banerjee
,
N.
,
Bove
,
C.P.
,
Cui
,
X.
,
Du
,
R.
,
Franklin
,
P.
,
Graff
,
J.P.
,
He
,
P.
,
Jaeger
,
A.
,
Li
,
L.
,
Berenson
,
D.
,
Gennert
,
M.A.
,
Feng
,
S.,
,
Liu
,
C.
,
Xinjilefu
,
X.
,
Kim
,
J.
,
Atkeson
,
C.G.
,
Long
,
X.
,
and
Padir
,
T.
,
“Team WPI-CMU: Achieving Reliable Humanoid Behavior in the DARPA Robotics Challenge”
,
J. Field Robotics
, Vol.
34
, pp.
381
399
,
2017
.
2.
Long
,
X.
,
Wonsick
,
M.
,
Dimitrov
,
V.
,
and
Padir
,
T.
,
“Anytime multi-task motion planning for humanoid robots,”
2017 IEEE/ RSJ International Conference on Intelligent Robots and Systems (IROS)
, Vancouver, BC, Canada, pp.
4452
4459
,
2017
.