Robonaut 2 – The First Humanoid Robot in Space

*M.A. Diftler, *N.A. Radford, *J.S. Mehling, **M.E. Abdallah, *L.B. Bridgwater, **A.M. Sanders, *R.S. Askew **D. M. Linn, ***J.D. Yamokoski, ***F.A. Permenter, *** B.K. Hargrave

*NASA/JSC , Houston, Texas **General Motors, Warren Michigan, ***Oceaneering Space Systems, Houston Texas

Abstract—NASA and General Motors have developed the second generation Robonaut, Robonaut 2 or R2, and it is scheduled to arrive on the International Space Station in late 2010 and undergo initial testing in early 2011. This state of the art, dexterous, anthropomorphic robotic torso has significant technical improvements over its predecessor making it a far more valuable tool for astronauts. Upgrades include: increased force sensing, greater range of motion, higher bandwidth and improved dexterity. R2’s integrated mechatronics design results in a more compact and robust distributed control system with a faction of the wiring of the original Robonaut. Modularity is prevalent throughout the hardware and software along with innovative and layered approaches for sensing and control. The most important aspects of the Robonaut philosophy are clearly present in this latest model’s ability to allow comfortable human interaction and in its design to perform significant work using the same hardware and interfaces used by people. The following describes the mechanisms, integrated electronics, control strategies and user interface that make R2 a promising addition to the Space Station and other environments where humanoid robots can assist people.

Robonaut 2 – The First Humanoid Robot in Space

*M.A. Diftler, *N.A. Radford, *J.S. Mehling, **M.E. Abdallah, *L.B. Bridgwater, **A.M. Sanders, *R.S. Askew **D. M. Linn, ***J.D. Yamokoski, ***F.A. Permenter, *** B.K. Hargrave *NASA/JSC , Houston, Texas **General Motors, Warren Michigan, ***Oceaneering Space Systems, Houston Texas

original Robonaut. Modularity Abstract—NASA and General is prevalent throughout the Motors have developed the hardware and software along second generation Robonaut, with innovative and layered Robonaut 2 or R2, and it is approaches for sensing and scheduled to arrive on the control. The most important International Space Station in aspects of the Robonaut late 2010 and undergo initial philosophy are clearly present testing in early 2011. This state in this latest model’s ability to of the art, dexterous, allow comfortable human anthropomorphic robotic torso interaction and in its design to has significant technical perform significant work using improvements over its the same hardware and predecessor making it a far interfaces used by people. The more valuable tool for following describes the astronauts. Upgrades include: mechanisms, integrated increased force sensing, electronics, control strategies greater range of motion, higher and user interface that make R2 bandwidth and improved a promising addition to the dexterity. R2’s integrated Space Station and other mechatronics design results in environments where humanoid a more compact and robust robots can assist people. distributed control system with a faction of the wiring of the

I. INTRODUCTION introduction of robots in GM manufacturing, many of the N ASA and General Motors have a history of working together, targeted applications for robot use taking on formidable challenges, remain the same. Current that date back to the Apollo Lunar industrial robots operate in a Rover (need reference). The two highly structured task environment organizations have come together and are designed and again and this time to address a programmed to work in enclosed new challenge, developing robot workcells. The consistency of the assistants that can work in task structure enables robots to proximity to humans. safely perform their tasks. (Brief paragraph about how we However, this same structure also respect Justin, Asimov, Partner limits the robot task flexibility. robots etc, with references) While there has been some General Motors has been a technical progress that enables leader in the application and robots to operate in manufacturing development of robotic technology operations with less structure, the since its initial collaboration with full technical capability is still not Joseph Engelberger. GM was the mature, and has not been first manufacturer to use industrial realized. This "capability gap" has robots with its application of limited the range of robot Unimate robots in 1961. Today, applications for less structured General Motors employs uses environments. over 25,000 robots in its Manufacturing Operations worldwide. GM has influenced the industry over the years by leading technical development efforts in servo electric welding robots, paint application robots, the Unimate PUMA robot for light assembly, and fixturing robots. Fig. 1: Robonaut 2: units A and B Although performance, capability and reliability have greatly NASA experiences a very similar improved since the first robotics “capability gap.” The

challenge in this case is to make are well suited for applications in more Extra-Vehicular Activity an unstructured environment and (EVA) tasks robotically are fully expected to reduce the compatible. Many of the work load on EVA crew by maintenance tasks on the performing routine maintenance, International Space Station (ISS) assisting crew members before, are robotically serviceable and the during, and after EVA, and serving Canadian Space Agency’s Special in a rapid response capacity. Purpose Dexterous Manipulator NASA’s success in developing (SPDM) currently on-board ISS and demonstrating R1’s has the capability to perform these capabilities attracted the attention tasks. However, to perform its job of GM. GM approached NASA in SPDM, utilizes different approach 2006 as part of a worldwide corridors than human EVA and review of humanoid robotics in must interface with specialized search of new technologies that robotically compatible interfaces. would help their skilled workforce While specialized worksites and improve product quality and interfaces for robotics system manufacturing assembly have been very successful in processes. A detailed review of space they only address a portion the Robonaut 1 system convinced of the servicing for ISS. GM that NASA’s expertise in Astronauts working with EVA upper body systems specifically compatible tools will perform a designed to assist astronauts considerable amount of the made the agency an excellent maintenance activities. partner for developing the robotic NASA developed Robonaut 1, technologies that would also meet R1, (ref) to assist the crew in GM’s goals of closing the these servicing tasks and reduce “capabilities gap”. It was realized this “capability gap.” R1 has by both organizations that there demonstrated, in high fidelity were enormous benefits from a ground based demonstrations, its robust anthropomorphic robotic ability to work with existing EVA system that relieves people - tools, and interfaces within the factory workers or astronauts - constraints of EVA worksites. from dangerous or ergonomically Anthropomorphic robots, like R1, painful and difficult activities.

For NASA and GM to achieve runs of sensitive analog signals. the desired performance Therefore, the avionic architecture improvements, Robonaut 2 for R2 was designed and required a number of significant developed around one central advancements in the robot's theme – the reduction of the electromechanical design, sensing conductor count in the robot and integration, controls strategy, and specifically the two arms. user interface. At the heart of In order to fundamentally reduce these advancements are the conductor count, a new technologies and approaches that communication scheme was allow for increased speed, required. R1 was built upon a strength, and dexterity while not point to point RS-485 sacrificing, and in fact improving communication architecture that, upon, a system design compatible including power and other relevant with direct human interaction that signals, required over 100 hads always been a focus of conductors in each main arm previous Robonaut development. cable. This produced an unwieldy (arkward) amount of wires that often got damaged during unrelated II. MECHATRONIC DESIGN FOR servicing and the wires RELIABILITY themselves were often the reason With 42 independent degrees of for other maintenance required on freedom, 50 motors and over 350 the robot. Moreover, that amount sensors, R2, shown in figure 1, is of wiring for each arm forced the a mechatronic integration cabling to reside externally challenge. Prior lab experience whereby it created additional on R1 and other robots challenges in how to manage to demonstrated a direct correlation service loop. In contrast, R2 was between the overall wire count in designed to have a distributed a robotic system and the reliability processing architecture with a of said system. R1 had a high speed serial communication centralized processing and point network that utilized a bussed to point communication paradigm power configuration. This afforded which necessitated large the robot a minimal set of wires for conductor count cabling and long each internal main arm cable

which totaled only 16. The high scale, 5 degree-of-freedom upper speed serial communication arms. The use of series elastic structure is a custom protocol actuation, however, differentiates which utilizes Multi-drop Low R2 from previous designs. Voltage Differential Signal Developed initially with legged (MLVDS) as the physical layer robots in mind, series elastic with bus speeds of 50Mbit. Each actuators (SEA's) have been node on the MLVDS network shown to provide improved shock makes up the vital string of tolerance, beneficial energy controllers that provide the low storage capacity, and a means for level joint control and data accurate and stable force control processing. The upper arm has 5 [1,2]. Not surprisingly, these nodes and the lower arm consists features are of interest to the of two additional nodes; one for manipulation community as well the forearm and one for the hand. and a number of humanoid robot arms have been designed with passive compliance between their joints' actuators and outputs [citation needed]. Robonaut 2 is unique among these robots, in that it does not sacrifice strength, or payload capacity, to achieve fine torque sensing at each of its joints. This is made possible by the custom planar torsion springs Fig. 2: A picture of multiple sizes that are integrated into each arm of arm spring actuator and the two 19 bit A. Upper Arms absolute angular position sensors Robonaut 2, like its predecessor, that measure each spring's uses brushless DC motors, deflection. Shown in Fig. 2, these harmonic drive gear reductions, springs are uniquely sized for and electromagnetic failsafe each arm joint and are capable of brakes as the building blocks for elastic deformation across the full the power and torque dense range of their actuators' actuators in the robot's human- continuous torques. This gives

Robonaut 2 the fine force PowerPC processor and resolution to implement a variety programmable logic coupled with of impedance control modes at a 3 phase brushless DC inverter both the joint and Cartesian level bridge. This Superdriver performs while still enabling the robot to motor commutation and current handle significant human-scale control, payloads as in Fig. 3. serialization/deserialization of joint data and commands, sensor processing, and control of the SEA’s noted above. The embedded processor run 10KHz sensor processing, current and SEA control loops. Programmable logic is used for the MLVDS and motor Fig. 3: A picture of R2 weight commutation, which can either be lifting (clearly seeing the 20lb) six-step or space vector. SEA control can be configured locally In order to fully realize the upper for joint torque or joint impedance. arm capability, a considerable The Superdriver provides a amount of capability had to be streamlined means of interfacing placed on the low level joint to the analog and digital joint controllers. In R1, the joint loops sensors. These sensors need were closed at a high level central only communicate with a local processor and this fundamentally processor which is responsible for limited the joint loop rates and the all the joint control. Sensors read type of sensors used in that at each joint include three motor control. Therefore, each upper phase currents, motor bus current, arm node consists of a highly temperature sensors, two integrated joint controller, called a absolute position sensors (APS), Superdriver, which serves as the and one incremental motor workhorse of the upper arm encoder. The motor current motion control strategy. The sensors are used for vector Superdriver is an FGPA based control and torque limiting. The controller with an embedded absolute position sensors are

used in tandem to measure SEA spring deflection. One APS is B. Hands used for joint position estimation. Robonaut 2’s hand, shown in The incremental motor encoder is Figure 4, is designed to more used for vector commutation and closely approximate human hand also gives a redundant capabilities than its predecessor, measurement of input SEA spring Robonaut (Lovchik paper position if harmonic drive reference). With a two-DOF wrist, dynamics are ignored. Finally, the and a 12-DOF hand with five output position is differentiated at fingers, the forearm houses 18 the joint to provide an estimate of motors, eight separate circuit joint velocity. boards. The forelimb is self- One of the main achievements of contained, requiring only power the Superdriver design is how and communication. Robonaut highly integrated the printed circuit 2’s hand has demonstrated its board is into the overall upper arm capabilities to use a large set of joint design. Each Superdriver EVA tools, conventional hand board plugs in to each joint with 2 tools, and soft goods. blind mate connectors. This The hand, shown in figure 5, is provides a clean and robust divided into a dexterous set used interface for each board and for manipulation and a grasping places zero stress on any related set used to maintain a stable wiring which prevents life and grasp while working with large fatigue related failures. Each tools. The dexterous set consists circuit board is also very of two three-DOF fingers, the accessible makinges routine index and middle finger, and a maintenance and repairs easy. four-DOF opposable thumb. The grasping set consists of two one- DOF fingers, the ring and little finger. All fingers are shock- mounted within the palm, giving the hand rugged grasping options.

Fig. 4: Robonaut Hand (real picture?)Design

Fig. 5: R2 Finger groupings Fig. 5: Robonaut 2 emulation of Cutkosky's Grasp Taxonomy The quality of the Robonaut 2 hand was determined by its ability Each digit is controlled by a set to emulate Cutkosky's grasp of linear actuators that are taxonomy[ref]. The original extrinsically located in the Robonaut hand was only able to forearm. An N+1 configuration successfully emulate ~50% of the [Salisbury ref] is implemented and grasps, while Robonaut 2 has the allows for a minimal set of ability to emulate 90% of the actuators for each DOF. Each grasps. Figure 6 shows the finger phalange has the ability to Robonaut 2 hand in the contain a six-DOF load cell (ref - Cutkosky's grasps. The increase patent number 7,784,363) for a in capability of the hand came total of 14 load cells. The sensors about by advancements made in provide all three axes of force and the thumb design. The Robonaut moment, allowing for 2 thumb has four DOFs, is the measurement of external contact same scale, and closely forces as well as shear force and approximates the kinematic layout slippage of objects held by of a human thumb. In addition to Robonaut 2. Additional sensors the kinematic layout, the thumb are embedded within the palm actuation scheme is designed providing tendon tension such that it is significantly stronger measurement. The tendon than the fingers that it is in tension sensors measure the opposition to. internal forces of the antagonistic actuation of each finger. The forearm contains all the

power/data distribution and III. CONTROL STRATEGY processing electronics for the A. Implementation finger and wrist actuators. Only 6 The control software for R2 is a wires pass from the upper arm to multi-threaded application spread the lower arm. There are 2 across two PowerPC processor different types of electronic control cards in a Compact PCI (cPCI) boards in the forearm. At the top chassis. The application code was level is the forearm controller and developed in ControlShellR and below that is the forearm driver runs on the VxWorksR operating electronics. A single forearm system. The first processor controller board serves as a handles all the collection and supervisor/ controller for the high-level processing of the forearm driver electronics and sensor information from the communicates via MLVDS with embedded network. It also the robot’s central IO processor. oversees an extensive safety There are 6 motor controller system evaluating the kinematics boards, each of which interface to and force levels in the system. and drive 3 finger or wrist motors. The second processor implements In total the motor controller boards the control law and computes the command 16 finger motors and 2 kinematics. Sensor data and wrist motors. computed quantities are shared The forearm motor driver boards between the two processors via a have phase current sensing and shared memory region through the provide the ability to set a PCI backplane. continuous current limit, as well as These central processors pass allow a brief current override in down desired torque commands case a pulse of current is required to the motor drivers. The for a short amount of time. A embedded processors there separate over-current bit will implement a high speed torque trigger if the max current threshold control loop running at 10 kHz. is exceeded. This triggers a fault in the driver which can be reset at B. Architecture the brainstem level once an The controls architecture was evaluation is performed. designed to provide an impedance based control system with great

flexibility. The concept of PID force control objective can impedance control provides for preempt the primary operational robust interaction with the space relation to create a mixed environment, while allowing for force/impedance control behavior both motion and force control in operational space. objectives (Hogan 1987). The The secondary joint-space architecture was designed to allow impedance relation then governs for multiple, prioritized tasks that the joint motion lying in the null- can control the robot with respect space of the primary task, to different spaces and different including the joints downstream of nodes of interest in the kinematic the selected node. When no node tree. At the heart of the on a respective arm is selected for framework is an impedance law the primary task, this joint-space with two hierarchical relations: an relation fully governs all joints of operational-space impedance at the arm. top priority, and a joint-space C. Arms Control Law impedance at second priority. The dual-priority control Multiple such laws can be described in the previous implemented hierarchically. subsection is defined as follows. The primary operational space First, consider the equations of relation can be defined with motion for the full system of respect to any of a number of manipulators. possible nodes in the tree, and it can be defined with respect to Mq+ c + g −τ = τ (1) either the linear and/or the angular e motion. Given the task definition, M is the joint-space inertia matrix. the relevant kinematics are q is the column matrix of joint generated in real-time for all joints angles. c is the column matrix of upstream of the selected node in Coriolis and centripetal the kinematic chain. The waist generalized forces. g is the joint can be optionally included in column matrix of gravitational the kinematic chain, allowing the generalized forces. And τ and τ arms to be controlled either e are the column matrices of independently or integrated for actuated and external torques, whole-body motion. In addition, a

τ = −J T (B x + K ∆x)− N(B q + K ∆q)+ g respectively. o o j j (3) Second, consider the desired N = I − J + J closed-loop impedance relations for both the operational and joint J is the Jacobian matrix mapping spaces. joint velocities to operational- space velocities, and N is the null- M x+ B x + K ∆x = F o o o e (2) space projection matrix for J. A M q + B q + K ∆q = τ j j j e closed-loop analysis shows that this control law provides the Mo, Bo, and Ko represent the desired joint-space impedance desired operational-space inertia, relation (2) projected into the null- damping, and stiffness matrices, space. In the range-space, it respectively. Mj, Bj, and Kj provides the desired operational- represent the desired joint-space space impedance relation (2) with inertia, damping, and stiffness a disturbance from the null-space matrices, respectively. x and Fe dynamics. The effects of this represents the operational-space disturbance, as well as the effects coordinates and corresponding of neglecting the Coriolis forces external forces, respectively. The and the derivative of J, are Δ indicates the error in the negligible at the speeds R2 respective variable with respect to operates at. its desired value. A similar relationship to (3) is To eliminate the need for used for mixed force/impedance sensing of the external torques, mode. A null-space projection the impedance inertias are set to matrix for the force task Jacobian the passive inertia of the system: is used to project the primary and -1 -1 T Mo = M, Mj = JM J . The full secondary tasks into the force solution for this dual-priority task’s null-space. impedance control law is D. Fingers Control Law presented in (Platt 2010). For the Since the fingers are actuated by sake of the implementation, the coupled tendons, rather than following approximation was independent drives, a special employed to eliminate the need control law is needed for them. It for the inertia matrix. turns out that the state-of-the art

in force control of tendon-driven

manipulators actually controlled emergency stop buttons as part of the manipulators in the tendon- their standard safety systems. In space, which exhibit a first-order the case of industrial manipulator, coupled disturbance between the additional safety devices such as joints. Alternatively, the control light curtains or sensor mats also law can be formulated in the joint- provide power cutoffs in the event space to eliminate this coupled of a human entering the robots disturbance and increase the workspace. R2 has moved away speed of the response. A full from this paradigm. The basic discussion of this control law is torque control strategy noted available in (Abdallah 2010). We above limits the force the robot present here the final control law, applies to the environment where an inner position control ensuring that when inadvertent loop is implemented on the contact does occur, the resulting actuators. force felt by the human is comfortable and the robot can be T  R  τ  manually restricted by pushing on    pc = p − kd p +   K p ∆  W   t  a limb. (4) In parallel to the basic torque control are software monitoring p and pc represent the actual and routines that use the torque commanded positions for the sensing at the arm and waist actuators. R (n × n+1) represents joints and multiple force sensors the kinematic mapping from in the arms, to independently tendon tensions to joint torques. monitor the robot’s forces. If a W is a row matrix selected predefined limit is exceeded at orthogonal to R. t is a scalar either the joint or the arm level, measure of the internal tension on the robot disengages motor power the tendon network, defined as t = and stops. A multitude of Wf. Kp and kd are constant gains. additional software routines on several processors continuously IV. WORKING IN PROXIMITY TO check the health and HUMANS communication of the two main Both industrial manipulators and force/torque monitoring loops. R1 utilize kill switches or The result is a triple redundant

system to keep forces within allowable limits. The benefit of this architecture is that the R2 on the ISS will not need an Emergency Stop button. This E-stop will be available and is now a Motion-Stop or M-stop. The robot operator has the discretion to use the M-stop as a convenience while using the The custom R2 interface uses robot, but crew will be allowed into common elements such as the robot’s workspace without buttons and sliders to provide the having to use it. user with direct control over the basic functionality of the robot. V. USER INTERFACE These objects are organized by One of the unique challenges of function and presented to the user working with R2 is distilling the based on the pre-determined control of a complicated system needs of a variety of tasks. This into a user-friendly design. In simple interface enables novice industrial and commercial users to perform basic tasks on environments, suppliers have the robot without requiring created interfaces to move, excessive training. monitor, and manage standard The interface uses a monitored robotic arms. Unfortunately, no network link to receive telemetry, commercial supplier offers a events, and status updates from standard interface to control a 42 the system and to send degree of freedom humanoid commands. The data is displayed robot. (expand to include teleop?) on the screen using information rich symbols and controls. The user interface modifies the visual appearance of these objects to highlight conditions that require additional attention. These objects and features enable the operator to filter the large amount

of available data and more easily actuate switches, use standard understand the state of the robot. tools, and manipulate softgood The interface also provides the objects and cables. The results of ability to log and graph both these experiments will help incoming and outgoing data. develop algorithms and control The interface is dynamically systems with even more configurable by the user. This capability. allows the operator to move the displays and controls to an VII. CONCLUSIONS orientation that is well suited for Robonaut 2 is a significant the current task. advancement in dexterous In addition to the basic control of robotics. Not only will it be a first the robot using regular input humanoid robot is space it methods, the user interface also represents a significant advance provides a custom program over its predecessor in terms of language. This language is strength, speed and dexterity. designed to work closely with the The… design methodologies ACKNOWLEDGMENT implemented in the robot The authors would like to thank hardware and controls algorithms. the entire Robonaut 2 team for The language provides a flexible their tireless dedication to way to integrate sequenced excellence: The team includes motion with the myriad of sensor the authors and the following platforms available to R2. individuals: Chris Ihrke (General VI. ACTIVITIES IN SPACE Motors), Don Davis (General The Robonaut 2 robot will be the Motors), Philip Strawser (NASA), first humanoid type robot sent to Adam Parson (OSS), Charles space. The robot will deployed on Wampler (General Motors), the International Space Station, Roland Menassa(General and will be used to perform a Motors), Matt Reiland (General variety of different experimental Motors)… (more to comeadded) tasks. While working side-by-side REFERENCES with its human astronauts, the R2 [1] Bluethmann, et al.l Robonaut robot will be tested for its ability to Astronaut Assistant.

[2] Variety of other humanoid Driven Fingers,” IEEE-RAS Intl. papers – due diligence Conf. on Humanoid Robots, [3] Pratt, G. A. and Williamson, M. Nashville, TN 2010. M. "Series Elastic Actuators." [9] R. Platt, M. Abdallah, C. in proceedings of IEEE/RSJ Wampler, “Multiple-priority International Conference on impedance control,” in Intelligent Robots and Systems submission to IEEE Intl. Conf. (IROS 95). Pittsburgh, PA. on Robotics and Automation, 1995. pp. 399-406. 2010. [4] Robinson, D. W. et al. "Series Elastic Actuator Development for a Biomimetic Walking Robot." in proceedings of IEEE/ASME International Conference on Advanced Intelligent Mechatronics. 1999. pp. 561-568 [5] Lovchik, C, Diftler, M. ICRA Robonaut paper 1999. [6] M. R. Cutkosky: “On Grasp Choice, Grasp Models, and the Design of Hands for Manufacturing Tasks”, IEEE Transactions on Robotics and Automation, Vol. 5, No. 3, pp. 269–279, 1989. [7] N. Hogan, “Impedance control: An approach to manipulation I- III,” Journal of Dynamic Systems, Measurement, and Control, vol. 107, pp. 1–24, 1987. [8] M. Abdallah, R. Platt, C. Wampler, B. Hargrave, “Applied Joint-Space Torque and Stiffness Control of Tendon-

4359847nnnm