0
Select Articles

From Torque-Controlled to Intrinsically Compliant Humanoid Robots PUBLIC ACCESS

[+] Author Notes
Christian Ott, Alexander Dietrich, Daniel Leidner, Alexander Werner, Johannes Englsberger, Bernd Henze, Sebastian Wolf, Maxime Chalon, Werner Friedl, Alexander Beyer, Oliver Eiberger, Alin Albu-Schäffer

German Aerospace, Center (DLR), Webling, Germany

1Corresponding author

Mechanical Engineering 137(06), S7-S11 (Jun 01, 2015) (5 pages) Paper No: ME-15-JUN-5; doi: 10.1115/1.2015-Jun-5

This paper gives an overview of the advancements in humanoid robotics at the German Aerospace Center (DLR) over the last 10 years. The development started with focus on dexterous, bimanual manipulation with the wheel-based humanoid Rollin’ Justin and continued with legged locomotion on TORO. With Rollin’ Justin, the team aims to create a cognitive robotic system that can reason about compliant manipulation tasks, based on intelligent decisions according to the actual state of the environment. These humanoids are expected to can perform a multitude of complex tasks and hereby contributing to human welfare. Possible fields of use include service robotics, industrial co-workers, search and rescue, space applications, medical robotics, etc. The experts suggest that teleoperated scenarios are feasible in short term, developing in long term towards shared or even full autonomy. Still, advancements must be made in almost all areas, starting from mechatronic robustness, reliability and energy efficiency, over multimodal perception and control up to autonomous planning and Artificial Intelligence-based reasoning. Development of interaction interfaces and communication modalities to humans will play an increasingly key role in the future.

For fulfilling predefined tasks of rather low complexity, specialized automats usually do a better job than general purpose machines, such as humanoid robots. Yet, for more complex and diverse tasks in a priori unknown human environments general purpose humanoids can provide a high degree of flexibility. In certain scenarios such as disaster management, robots may have to use tools and navigate through environments that were designed for humans. This is obviously true also if we think about future general purpose household robots, able to carry out all human housekeeping tasks. This motivates the design of anthropomorphic robots such as TORO and Rollin’ Justin. Also in tele-operation scenarios, human-like robots tend to be more intuitively operated by humans, due to their kinematic similarity.

There is one further strong reason why researchers build humanoid robots: understanding and technically reproducing such seemingly simple human tasks like dexterous grasping and manipulation, balancing, walking and running, perceiving the surrounding environment for planning and executing daily tasks are still largely unsolved questions, at least when compared to the human performance. Thus humanoid robot research helps to answer fascinating questions about human capabilities on the one hand, provides clues to build more dexterous, efficient and general purpose machines on the other hand.

In this paper we give an overview of the advancements in humanoid robotics at the German Aerospace Center (DLR) over the last decade.

The development started with focus on dexterous, bimanual manipulation with the wheel-based humanoid Rollin’ Justin and continued with legged locomotion on TORO. Both robots are characterized by torque-controlled actuators, capable of emulating the adaptable human muscle compliance by feedback control. A new generation of actuators is developed for the humanoid upper body HASY (Hand Arm System), in which “muscle compliance” is realized mechanically, by variable compliance actuators. This step promises increased impact robustness and energy efficiency by elastic energy storage, but raises at the same time substantial additional challenges regarding mechatronic integration and control.

The DLR Light-Weight-Robot-III [1] represents the third generation of torque-controlled robot arms developed at DLR. One of the main features of this robot is a tight mechatronic integration of strain-gauge-based torque sensors at the power-output side of the drive units. Such torque sensors allow for effective vibration damping in highly dynamic operations. Moreover, low-level torque feedback loops produce a highly sensitive back-drivable closed-loop behavior despite the highly geared drive units required by the lightweight construction of the joints. As a result, model-based nonlinear control approaches, such as impedance control, can be implemented successfully based on this technology. These drive units build a mature technological basis for the complex humanoid robots Rollin’ Justin and TORO (Figure 1).

Figure 1 Robots described in this article: Torque-controlled humanoid robots TORO and Rollin’ Justin, and the elastic Hand Arm System.

Grahic Jump LocationFigure 1 Robots described in this article: Torque-controlled humanoid robots TORO and Rollin’ Justin, and the elastic Hand Arm System.

Autonomous Compliant Manipulation with Rollin’ Justin

The mobile humanoid robot Rollin’ Justin is utilized as a research platform for autonomous planning and control of manipulation tasks in human environments. The system consists of an omnidirectional platform, an articulated torso, two seven degrees-of-freedom arms, two four-fingered dexterous hands, and a multi-sensory head (see Table 1 & Table 2). The hands are equipped with position and torque sensors and can thus be used for complex manipulation tasks: for example for handling tools or unscrewing the lid of a container or bottle. Rollin’ Justin can be operated without wires for about one hour. The size and geometry of the footprint of the mobile base can be adapted to the task by coordinating the movements of the four steerable spring-borne wheels. Overall, the robot can reach objects up to a height of 2.7m while still fitting through standard doorways. The vision system consisting of RGB-D cameras mounted in the head and the platform and a stereo camera pair allows for the 3D reconstruction of the environment.

With Rollin’ Justin we aim to create a cognitive robotic system that is able to reason about compliant manipulation tasks, based on intelligent decisions according to the actual state of the environment. In order to cope with a wide variety of tasks, we utilize a knowledge-based hybrid reasoning system to plan the task execution autonomously on the symbolic level (i.e. which actions have to be scheduled to satisfy the commanded goal state) and on the geometric level (i.e. what are appropriate task parameters to manipulate the objects involved in the actions) [2]. Moreover, during the reasoning procedure, the robot parametrizes the control level for each task execution individually. A hierarchical whole-body impedance control framework [3] builds the behavioral basis for the higher-level reasoning system (Figure 2). For each task, the robot selects and parameterizes the required control strategies (e.g. Cartesian impedance control, singularity avoidance, and self-collision avoidance) and the controller parameters (i.e. Cartesian trajectory, Cartesian stiffness, and maximum allowed Cartesian forces) based on the requirements of the objects involved in the task execution and the environment.

Figure 2 Rollin’ Justin combines a multi-task whole-body impedance controller with a high level reasoning unit acting both on symbolic, object and task level, and on the geometric level of the motion planner.

Grahic Jump LocationFigure 2 Rollin’ Justin combines a multi-task whole-body impedance controller with a high level reasoning unit acting both on symbolic, object and task level, and on the geometric level of the motion planner.

Exemplary tasks in domestic environments involve wiping of windows, cleaning the dishes and collecting dust or shards with a broom, as demonstrated in the video [4]. These tasks share the need for coordinated whole-body motions, while a tool is guided along a surface in contact. The tasks can be executed with the same overall control strategy, only requiring a different parameterization.

Balancing and Walking with TORO

While Rollin’ Justin's main focus is on safe human-robot interaction, complex whole-body motions, bimanual manipulation and other high-level tasks, the bipedal humanoid TORO was built with the aim of evaluating similar torque-based control concepts also for a legged robot. The relevant tasks for TORO include bipedal walking and multi-contact balancing, i.e. compliant stabilization against external disturbances while sustaining two or more end-effectors in contact. In contrast to the dexterous torque-controlled hands of Rollin’ Justin, the hands of TORO are human hand prostheses (iLimb ultra) allowing for a robust grip in multi-contact operations but without sensor feedback. Six-axis force-torque sensors in the feet allow measurement of the Zero-Moment-Point (ZMP), i.e. the torque-free point of action of the gross contact force, an inertia measurement unit (IMU) in the trunk is used for real-time control. In accordance with the aim of studying dynamic walking approaches, the feet were designed relatively small, having a size of 19 x 9,5cm. The multi-sensory head consists of a stereo camera, a RGB-D sensor and an additional IMU, which are fused by an onboard computer to provide an ego-motion estimation (based on an extended Kalman filter) and a mapping of the environment. The onboard batteries in the backpack allow for an autonomous operation of up to 1h.

Our approach for the generation and stabilization of walking motions is based on the concept of Capture Point, which is defined as the point on the floor where the robot has to place the ZMP in order to stop within one step. It can be shown that the use of the Capture Point as a state variable separates the overall dynamics into the stable and unstable part. For gait stabilization we utilize an underlying position-based ZMP controller and treat the ZMP as the control input. Moreover, from the Capture Point dynamics one can also see that a sequence of constant ZMP locations (associated with the footsteps) leads to a Capture Point trajectory, which geometrically is simply a connection of lines (zig-zag-curve, Figure 3). As a consequence, the trajectory generation can be performed in a highly efficient way as part of the real-time controller [5].

Figure 3 Illustration of the control approaches used for the bipedal humanoid TORO. Planning of constant reference ZMP locations in the feet imply a linear evolution of the reference Capture Point during walking. In situations involving multiple contact points, desired control forces on the CoM are distributed amongst the available contacts.

Grahic Jump LocationFigure 3 Illustration of the control approaches used for the bipedal humanoid TORO. Planning of constant reference ZMP locations in the feet imply a linear evolution of the reference Capture Point during walking. In situations involving multiple contact points, desired control forces on the CoM are distributed amongst the available contacts.

Motivated by the successful implementation of torque-based impedance controllers for manipulation with Rollin’ Justin, we developed a balancing controller for TORO which builds up on the torque-controlled operation mode of the joint drive units. The controller aims at generating a desired wrench (6-dimensional force-torque vector) at the CoM of the robot [6]. This desired wrench contains a compensation of the robot's total gravity force and a proportional and derivative control action responding to deviations of the CoM and hip orientation from a desired equilibrium configuration. Then a set of contact forces for the end-effectors in contact is computed by an optimization formulation considering unilaterality and friction cone constraints. Finally, the contact forces for all end-effectors are realized by mapping the forces into desired joint torques, which are transferred as set-points to the underlying joint torque controllers. The algorithm has been evaluated in a series of balancing experiments with two (only feet) to four (feet and hands) end-effectors in contact (see Figure 3), including balancing on movable inclined planes, rocks, and even on compliant surfaces (sports mattresses). Current extensions of this controller focus on the realization of dynamic changes in the number of contacts as well as on combinations with the Capture-Point-based algorithm for gait stabilization.

What does it aim at?

The Hand Arm System is a DLR development towards the next generation of humanoid robots in terms of mechatronic design. The aim is to reach the performance of human beings in terms of speed, force and accuracy [7]. Its design philosophy is to understand the biological system and implement the technology to provide a functional equivalent but avoid making a blind copy of the biology.

How does it work?

In humans, the elasticity provided by the muscles, tendons and ligaments decouples the link position from the drive position. Generally speaking, the energy introduced into the system, no matter whether caused by a collision, external forces or acceleration of the link inertia, is converted to elastic energy. This power source can be used to regain kinetic energy and therefore enhances the dynamics of the system. This motivated the introduction of mechanical springs, placed between the output of the gear box and the link to provide a similar behavior. Moreover, by using several nonlinear mechanisms actuated by two motors per joint, it is possible to adjust the stiffness of the joints and adapt to the task requirements.

System Overview

The Hand Arm System is an upper body humanoid robot with two arms and hands. All of its 48 joints are actuated with nonlinear, adjustable stiffness mechanisms. It is equipped with more than 300 sensors and 100 motors that are controlled at a frequency of 3kHz. We experimented with different concepts of implementing variable intrinsic compliance [8].

Current Work

The platform is used to investigate and experiment modeling and control but also on new planning and grasping strategies (Figure 4).

Figure 4 Execution of a plan obtained by grasp and arm planning. In-hand object localization is obtained by fusion of kinematic, tactile and vision data. A particle filter uses a simplified object model and the robot kinematics and tactile sensing capabilities in order to discard or promote object location hypothesis. On the left, a typical pick-and-place task is executed. On the right top, the picture depicts an invalid hypothesis (collision and unexplained contact). On the right bottom, the diagram depicts how the particles reflect the hypothesis quality (the larger the circle, the better the hypotheses). The approach allows monitoring the grasp execution interactively and significantly improves the success rates.

Grahic Jump LocationFigure 4 Execution of a plan obtained by grasp and arm planning. In-hand object localization is obtained by fusion of kinematic, tactile and vision data. A particle filter uses a simplified object model and the robot kinematics and tactile sensing capabilities in order to discard or promote object location hypothesis. On the left, a typical pick-and-place task is executed. On the right top, the picture depicts an invalid hypothesis (collision and unexplained contact). On the right bottom, the diagram depicts how the particles reflect the hypothesis quality (the larger the circle, the better the hypotheses). The approach allows monitoring the grasp execution interactively and significantly improves the success rates.

A typical application demonstrating the potential of compliant actuators is illustrated in Figure 1 (bottom). The arm is driven in mechanical resonance to achieve link velocities above the motor velocity and allow impact torques which are above the maximum motor and gearbox torques. The impact force peak is absorbed and smoothened by the spring. Despite the large actuator compliance, positioning precision is achieved by iterative learning control.

A disadvantage of the very compliant actuation is the low damping of the system when performing fast positioning motions based only on motor position information. However, measuring the joint torque and its derivative based on the spring deflection [9] allows applying nonlinear control techniques to effectively damp out these oscillations (Figure 5).

Figure 5 Improvement of trajectory tracking by active vibration damping (bottom).

Grahic Jump LocationFigure 5 Improvement of trajectory tracking by active vibration damping (bottom).

Our overall goal is to develop safe and robust humanoid robots that are capable of performing a multitude of complex tasks and hereby contributing to human welfare. While a decade ago, humanoids seemed far too complex for realistic scenarios, the current results encourage us to imagine first applications within the next decade. Possible fields of use include service robotics, industrial coworkers, search and rescue, space applications and medical robotics, to name but a few. Teleoperated scenarios are feasible in short term, developing in long term towards shared or even full autonomy. Still, advancements have to be made in almost all areas, starting from mechatronic robustness, reliability and energy efficiency, over multimodal perception and control up to autonomous planning and AI-based reasoning. Development of interaction interfaces and communication modalities to humans will play an increasingly important role in the future.

Albu-Schaeffer, A., Haddadin, S., Ott, Ch., Stemmer, A., Wimboeck, T., and Hirzinger, G., 2007, The DLR Lightweight Robot-Design and Control Concepts for Robots in Human Environments, Industrial Robot: An International Journal, 34 (5), pp. 376– 385 DOI: 10.1108/01439910710774386 [CrossRef]
Leidner, D., Borst, Ch., and Hirzinger, G., 2012, “Things are made for what they are: Solving manipulation tasks by using functional object classes”, 12th IEEE-RAS International Conference on Humanoid Robotics, pp. 429-435, 2012. DOI:10.1109/HUMAN0IDS.2012.6651555
Dietrich, A., Wimboeck, T., Albu-Schaeffer, A., and Hirzinger, G., 2012, Reactive Whole-Body Control: Dynamic Mobile Manipulation Using a Large Number of Actuated Degrees of Freedom1, IEEE Robotics & Automation Magazine (RAM), 19 (2), pp. 20– 33 DOI:10.1109/MRA. 2012.2191432 [CrossRef]
Leidner, D., and Dietrich, A., 2015, “Towards Intelligent Compliant Service Robots”’, Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, TX.
Englsberger, J., Ott, Ch., and Albu-Schäffer, A., 2015, Three-dimensional bipedal walking control based on Divergent Component of Motion, IEEE Transactions on Robotics (TRO), 31 (2), pp. 355– 368 DOI: 10.1109/ TR0. 2015.2405592 [CrossRef]
Ott, Ch., Roa, M. A., and Hirzinger, G., 2011 “Posture and Balance Control for Biped Robots based on Contact Force Optimization”, 11th IEEE-RAS International Conference on Humanoid Robots, pp. 26-33, Bled, Slovenia. DOI:10. 1109/ Humanoids. 2011. 6100882
Grebenstein, M., et al. , 2011 The DLR Hand-Arm System, IEEE International Conference of Robotics and Automation (ICRA), pp. 3175– 3182. DOI:10.1109/ICRA. 2011.5980371
Wolf S., et. al., 2015, “Soft Robotics with Variable Stiffness Actuators: Tough Robots for Soft Human Robot Interaction”, Soft Robotics, Springer Verlag, pp 231-254. D0I:10.1007/978-3-662-44506-8_20
Petit, F., and Albu-Schaeffer, A., 2011, State Feedback Damping Control For A Multi DOF Variable Stiffness Robot Arm, IEEE International Conference on Robotics and Automation, pp. 5561– 5567 DOI: 10.1109/ICRA.2011.5980207
Copyright © 2015 by ASME
View article in PDF format.

References

Albu-Schaeffer, A., Haddadin, S., Ott, Ch., Stemmer, A., Wimboeck, T., and Hirzinger, G., 2007, The DLR Lightweight Robot-Design and Control Concepts for Robots in Human Environments, Industrial Robot: An International Journal, 34 (5), pp. 376– 385 DOI: 10.1108/01439910710774386 [CrossRef]
Leidner, D., Borst, Ch., and Hirzinger, G., 2012, “Things are made for what they are: Solving manipulation tasks by using functional object classes”, 12th IEEE-RAS International Conference on Humanoid Robotics, pp. 429-435, 2012. DOI:10.1109/HUMAN0IDS.2012.6651555
Dietrich, A., Wimboeck, T., Albu-Schaeffer, A., and Hirzinger, G., 2012, Reactive Whole-Body Control: Dynamic Mobile Manipulation Using a Large Number of Actuated Degrees of Freedom1, IEEE Robotics & Automation Magazine (RAM), 19 (2), pp. 20– 33 DOI:10.1109/MRA. 2012.2191432 [CrossRef]
Leidner, D., and Dietrich, A., 2015, “Towards Intelligent Compliant Service Robots”’, Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, TX.
Englsberger, J., Ott, Ch., and Albu-Schäffer, A., 2015, Three-dimensional bipedal walking control based on Divergent Component of Motion, IEEE Transactions on Robotics (TRO), 31 (2), pp. 355– 368 DOI: 10.1109/ TR0. 2015.2405592 [CrossRef]
Ott, Ch., Roa, M. A., and Hirzinger, G., 2011 “Posture and Balance Control for Biped Robots based on Contact Force Optimization”, 11th IEEE-RAS International Conference on Humanoid Robots, pp. 26-33, Bled, Slovenia. DOI:10. 1109/ Humanoids. 2011. 6100882
Grebenstein, M., et al. , 2011 The DLR Hand-Arm System, IEEE International Conference of Robotics and Automation (ICRA), pp. 3175– 3182. DOI:10.1109/ICRA. 2011.5980371
Wolf S., et. al., 2015, “Soft Robotics with Variable Stiffness Actuators: Tough Robots for Soft Human Robot Interaction”, Soft Robotics, Springer Verlag, pp 231-254. D0I:10.1007/978-3-662-44506-8_20
Petit, F., and Albu-Schaeffer, A., 2011, State Feedback Damping Control For A Multi DOF Variable Stiffness Robot Arm, IEEE International Conference on Robotics and Automation, pp. 5561– 5567 DOI: 10.1109/ICRA.2011.5980207

Figures

Tables

Table Grahic Jump Location
Table 1 Overview of the main characteristics of the humanoid systems described in this article.
Table Grahic Jump Location
Table 2 Degrees of freedom of the three humanoid systems described in this article.

Errata

Discussions

Some tools below are only available to our subscribers or users with an online account.

Related Content

Customize your page view by dragging and repositioning the boxes below.

Related Journal Articles
Related eBook Content
Topic Collections

Sorry! You do not have access to this content. For assistance or to subscribe, please contact us:

  • TELEPHONE: 1-800-843-2763 (Toll-free in the USA)
  • EMAIL: asmedigitalcollection@asme.org
Sign In