Our Robots

At IAS, we have access to a series of really great robots:

  1. DarIAS: Bimanual manipulation platform
  2. Iiwas: Floor-mounted bimanual manipulation platform
  3. KoBo: Assistive bimanual manipulation platform
  4. Barrett WAM: High-speed Robot for dynamic tasks
  5. Quanser Platforms: Physical RL benchmark systems
    1. Furuta Pendulum
    2. Cart-Pole
    3. Magnetic Levitation
    4. Ball-Balancer
    5. Omni Bundle
  6. BioRobs: Compliant arms for Tetherball
  7. Robot Table Tennis: Two opposing Barrett WAMs for playing table-tennis
  8. iCub
  9. Nao
  10. Mitsubishi PA-10
  11. Allegro hand with BioTacs
  12. Wessling Robotic Hands with BioTacs
  13. aDDa4students Car

Darias: our Bimanual Manipulation Platform

The robot Darias (DARmstadt IAS) is our main platform for research into bimanual and dexterous manipulation. The robot consists of a torso with two Kuka light weight robot arms, each of which has a five-fingered DLR hands as an end effector. For observing its envrionment, the robot is equipped with a Kinect and connected with our Optitrak system. The optitrak allows for marker-based tracking of objects and humans at a rate of 90Hz.

Each arm has seven degrees of freedom in an anthropomorphic configuration, i.e., three shoulder joints, an elbow, and three wrist joints. Communication with the robot runs at 1kHz, and allows for torque control of the robot's joints. The robot's joints are equipped with torque sensors as well as joint encoders. The robot arms are actively compliant, which allows them to be easily used for kinaesthetic teachin. The active compliance helps the robot to safely interact with its environment and with humans.

The five-fingered hands of the robot also have an anthropomorphic design. Each finger has three active degrees of freedom, including proximal and distal joints for flexing and extending the fingers, as well as a third joint in the base that allows the robot to spread its fingers apart. Similar to the robot arm, the joints of the robot's fingers provide torque information as well as the joint angle. The fingers are controlled using joint impedance control, which makes them actively compliant. This compliance of the finger, as well as the arms, allows the robot to better handle uncertainty in its surroundings .

Some of our work with Darias:

  1. Maeda, G.; Ewerton, M.; Osa, T.; Busch, B.; Peters, J. (2017). Active Incremental Learning of Robot Movement Primitives, Proceedings of the Conference on Robot Learning (CoRL).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  2. Maeda, G.; Neumann, G.; Ewerton, M.; Lioutikov, R.; Kroemer, O.; Peters, J. (2017). Probabilistic Movement Primitives for Coordination of Multiple Human-Robot Collaborative Tasks, Autonomous Robots (AURO), 41, 3, pp.593-612.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  3. Lioutikov, R.; Maeda, G.; Veiga, F.F.; Kersting, K.; Peters, J. (2020). Learning Attribute Grammars for Movement Primitive Sequencing, International Journal of Robotics Research (IJRR), 39, 1, pp.21-38.   See Details [Details]   BibTeX Reference [BibTex]

Interested in this robot system? Please contact Julen Urain De Jesus!

Iiwas: Our Floor-Mounted Bimanual Manipulation Platform

We also have a floor-mounted bimanual manipulation platform based on two Kuka LBR R820 manipulators. The manipulators are similar to the LWR 4+ used by DarIAS and are impedance-controlled at a rate of 1kHz. Both arms a equipped with a SAKE Gripper and an Intel RealSense D435 RGB-D camera. The robots are mounted together with our Mitsubishi PA-10 on a common frame that allows for flexible arrangements. We use six Optitrack Flex 13 cameras for marker-based tracking.

Some of our work with Iiwas:

  1. Wibranek, B.; Belousov, B.; Sadybakasov, A.; Tessmann, O. (2019). Interactive Assemblies: Man-Machine Collaboration through Building Components for As-Built Digital Models, Computer-Aided Architectural Design Futures (CAAD Futures).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Interested in this robot system? Please contact Oleg Arenz!

Kobo: Our Assistive Bimanual Manipulation Platform

We also have a another bimanual manipulation platform, consisting of two FRANKA PANDA arms and a pan tilt camera head. Both arms a equipped with a FRANKA Hand and a Intel RealSense D435 RGB-D camera is mounted on the pan-tilt head. The full system is integrated in ROS which allows for easy integration of additional sensors, such as an eyetracker or microphones.

The KoBo cell is additionally equipped with a motion tracking system where use six Optitrack Flex 13 cameras for marker-based tracking.

Interested in this robot system? Please contact Dorothea Koert!

High-speed Barrett WAM

An exclusive high-speed 7 degrees-of-freedom version of the famous Barrett WAM robot has recently arrived at our lab in Darmstadt. This cable driven robot is capable of producing extremely high accelerations and is uniquely suited for studying highly dynamic movements that lie beyond the capabilities of standard industrial robots. Our low-level torque control interface tightly integrated with a simulation environment as well as with an OptiTrack object tracking system allows for fast prototyping and rapid experimentation with the robot. A Robcom interface makes it easy to use a familiar language and environment such as Python, Matlab, or ROS for quickly testing new algorithmic ideas.

Several ongoing projects—including badminton, beerpong, and juggling—provide a great opportunity for motivated students to learn more about real-time control of nonlinear dynamical systems, as well as apply their knowledge of robot learning and machine learning in challenging control problems.

Some of our work with Barrett WAM:

  1. Nass, D.; Belousov, B.; Peters, J. (2019). Entropic Risk Measure in Policy Search, Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  2. Lutter, M.; Ritter, C.; Peters, J. (2019). Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning, International Conference on Learning Representations (ICLR).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  3. Brandherm, F.; Peters, J.; Neumann, G.; Akrour, R. (2019). Learning Replanning Policies with Direct Policy Search, IEEE Robotics and Automation Letters (RA-L).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Interested in this robot system? Please contact Boris Belousov!

Quanser Platforms

The Quanser platforms are a set of control systems for educational purposes. In our lab, we have a Furuta pendulum, a single and a double Cart-Pole, a Magnetic Levitation, a 2DOF Ball Balancer and a haptic-device called Omni Bundle. Most of the system are classical control tasks, often used as benchmarks in optimal control and reinforcement learning.

We provide a python library for an easy interaction with the platform, using an interface similar to OpenAI-Gym.

Furuta Pendulum

The Furuta-Pendulum is a rotational inverted pendulum, an under-actuated invented by Katsuhisa Furuta and colleagues at Tokyo Institute of Technology in 1992. Since then, it has become a standard research platform for demonstrating performance of linear and non-linear control laws. Robust hardware implementation by Quanser makes this platform ideal for experimentation with reinforcement learning algorithms aimed at controlling non-linear physical systems. The prototypical task is swing-up and stabilization; however, other tasks including underactuated manipulation of external objects can be defined.

Some of our work with the Furuta Pendulum:

  1. Lutter, M.; Peters, J. (2019). Deep Lagrangian Networks for end-to-end learning of energy-based control for under-actuated systems, International Conference on Intelligent Robots and Systems (IROS).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Interested in this platforms? Please contact Boris Belousov!


The cart-pole is perhaps one of the most well-known control system used in control theory and in reinforcement learning. It consists of an actuated cart which can freely move on a horizontal track. It has attached one link (or two in the case of the double pendulum). Often the task consists in controlling the cart in such the way to swing the pendulum and eventually stabilize it in the up-right position.

Interested in this platforms? Please contact Samuele Tosatto or Michael Lutter!

Magnetic Levitation

This platform is a classical control experiment that miniturizes the Transrapid Levittion Train. The goal is to control and stabilize a metal ball via a magentic field. The input of the system is the current passing through the coil which modulates the magnetic field. The system is highly nonlinear, but it is usually linearized around an operating point and controlled with a cascade loop. Replacing this simple controller with a nonlinear one in a reinforcement learning setup is a challenging task.

Interested in this platforms? Please contact Hany Abdulsamad!


The Ball-Balancer can be classified as nonlinear under-actuated balancing problem with continuous 8D state and 2D action spaces. The default task is to stabilize the randomly initialized ball at the plate’s center. Given measurements and their first derivatives of the motor shaft angles as well as the ball’s position relative to the plate’s center, the agent controls two servo motors via voltage commands. The rotation of the motor shafts leads, through a kinematic chain, to a change in the plate angles. Finally, the plate’s deflection gets the ball rolling.

Interested in this platforms? Please contact Fabio Muratore!

Omni Bundle

The Omni Bundle is an haptic device with six revolute joints, and a stylus which you can use to move the robot. Default simulations include an environment where you can move some blocks, and a table tennis simulator where you can bounce a ball. The robot is more oriented to human-robot interaction than to control compared to the other platforms, but it is still possible to implement simple controllers and tasks (such as "goto" or "follow a trajectory" tasks).

Interested in this platforms? Please contact Simone Parisi!

Three BioRob compliant robotic arms

The BioRob arm is a compliant robotic arm which, depending on the version, has five or six degrees of freedom. It's tendon driven design kinematically decouples the joint and motor side and allows the heavy servo motors to be placed close to the base, the ``torso'', of the robot. The result is a super lightweight design, especially at the final links of the robot, that offers significant advantages for dynamic and high-speed movements. Additionally, the use of springs to connect the tendons provide compliance, a necessary property for striking movements such as hammering, and allow the storage and release of energy to gain even higher accelerations than the motors can provide. Overall BioRob's lightweight design offers a great platform for high-speed movements while minimizing the risk of damaging it's servo motors and increase safety even for close human-robot interaction.

But these advantages of the design come at a cost: controlling the robot is a complex problem that requires sophisticated control policies. At IAS we focus on improving the control performance of the robot on motor skill tasks. We generate novel model-based control approaches for controlling the robots which take into account the elasticity and the spring characteristic of the robots. Since creating models based just on CAD data lead to inferior performance, we use model learning approaches to improve the models. Additionally we use imitation learning for incorporating expert knowledge in our control policies and we subsequently improve the policies with reinforcement learning techniques. We evaluated the performance of our control approaches on hitting static and moving balls, but we also developed a two-robot setup, in which the robots compete on the game of tether-ball, for further experimentation.

Some of our work with BioRob:

  1. Kollegger, G.; Ewerton, M.; Wiemeyer, J.; Peters, J. (2017). BIMROB -- Bidirectional Interaction Between Human and Robot for the Learning of Movements, in: Lames, M.; Saupe, D.; Wiemeyer, J. (eds.), Proceedings of the 11th International Symposium on Computer Science in Sport (IACSS 2017), pp.151--163, Springer International Publishing.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  2. Parisi, S.; Abdulsamad, H.; Paraschos, A.; Daniel, C.; Peters, J. (2015). Reinforcement Learning vs Human Programming in Tetherball Robot Games, Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  3. Englert, P.; Paraschos, A.; Peters, J.;Deisenroth, M.P. (2013). Probabilistic Model-based Imitation Learning, Adaptive Behavior Journal, 21, pp.388-403.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Interested in these robot systems? Please contact Simone Parisi and Marco Ewerton!

Robot table tennis setup consisting of Barrett WAM and high-speed cameras

We have setup a highly advanced robot table tennis setup consisting of a high-speed, high voltage special-made version of the Barrett WAM robot together with eight high-speed Prosilica Cameras. The WAM is torque controlled at 500 Hz via CAN Bus and, due to the special make, can start nearly instantaneous to high accelerations. The Prosilica Cameras are operated at 200 Hz and are being used with our vision system described in Lampert, C.H.; Peters, J. (2012). Real-Time Detection of Colored Objects In Multiple Camera Streams With Off-the-Shelf Hardware Components, Journal of Real-Time Image Processing, 7, 1, pp.31-41.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex] .

The whole setup is located at our Tuebingen Lab location, the Robot Learning Lab at the Department for Empirical Inference at the Max Planck Institute for Intelligent Systems. Here, we have both students and post-docs, and many of our current members in Darmstadt have spend significant time at Tuebingen. We have used this setup for a series of motor skill learning tasks including Ball-in-a-Cup, Ball-Paddling and basic Robot Table Tennis.

Some of our work with the Table Tennis System:

  1. Daniel, C.; Neumann, G.; Kroemer, O.; Peters, J. (2016). Hierarchical Relative Entropy Policy Search, Journal of Machine Learning Research (JMLR), 17, pp.1-50.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  2. Muelling, K.; Kober, J.; Kroemer, O.; Peters, J. (2013). Learning to Select and Generalize Striking Movements in Robot Table Tennis, International Journal of Robotics Research (IJRR), 32, 3, pp.263-279.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  3. Daniel, C.; Neumann, G.; Peters, J. (2012). Hierarchical Relative Entropy Policy Search, Proceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS 2012).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  4. Kober, J.; Peters, J. (2010). Imitation and Reinforcement Learning - Practical Algorithms for Motor Primitive Learning in Robotics, IEEE Robotics and Automation Magazine, 17, 2, pp.55-62.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Interested in this robot system? Please contact Jan Peters!

iCub humanoid with skin and Coman legs

We have a full humanoid iCub robot (53 DOF), equipped with actuated cameras for stereo-vision, inertial sensor, whole-body skin (arms, legs, torso and foot-sole), tactile elements on the fingertips, 6 axis force/torque sensors (arms and legs), and variable-impedance actuation in the legs (design inherited from Coman's legs). Our version is the state-of-the-art and the best configuration for whole-body motions with contacts, such as walking or getting up from a chair. It is also the best configuration for physical interaction with humans and environment.

The setup is located in TU Darmstadt's Lab, where iCub has been used for the CoDyCo project.

Some of our work with iCub:

  1. Rueckert, E.; Camernik, J.; Peters, J.; Babic, J. (2016). Probabilistic Movement Models Show that Postural Control Precedes and Predicts Volitional Motor Control, Nature PG: Scientific Reports, 6, 28455.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  2. Calandra, R.; Ivaldi, S.; Deisenroth, M.;Rueckert, E.; Peters, J. (2015). Learning Inverse Dynamics Models with Contacts, Proceedings of the International Conference on Robotics and Automation (ICRA).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  3. Ivaldi, S.; Nguyen, S.M.; Lyubova, N.; Droniou, A.; Padois, V.; Filliat, D.; Oudeyer, P.-Y.; Sigaud, O. (2014). Object learning through active exploration, IEEE Transactions on Autonomous Mental Development, 6, pp.56-72.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Interested in this robot system? Please contact Svenja Stark!

Aldebaran Nao

The Aldebaran Nao (v4) is a 25 degrees of freedom humanoid. Its shape enables him to move and adapt to the world around him. His inertial unit enables him to maintain his balance and to know whether he is standing up or lying down. This robot has numerous sensors in his head, hands, and feet, as well as his sonars, enable him to perceive his environment and get his bearings. NAO is equipped with two cameras that film his environment in high resolution, helping him to recognize shapes and objects. With his 4 directional microphones and loudspeakers, Nao interacts with humans in a completely natural manner, by listening and speaking. It is possible to connect to the robot both using the Ethernet port and the WiFi connection. Nao is a platform easy to use and can be considered as the first choice for testing the learning of motion primitives. Given that the robot is already capable of replicating some human-like behavior, it is also possible to use this platform to learn complex behaviors, such as human-robot interaction tasks.

Interested in this robot system? Please contact Davide Tateo!

Mitsubishi PA-10

Our Mitsubishi PA-10 robot is a typical industrial robot arm with seven degrees of freedom. It has an internal PD controller with high gains, so it is position controlled. In the past, we have equipped the robot with different kinds of sensors and actuators, such as a force-torque sensor, an RGBD camera, and different kinds of tactile sensors. The PA-10 robot arm is mainly in use by the grasping and manipulation lab.

Some of our work with the PA 10:

  1. van Hoof, H.; Kroemer, O; Peters, J. (2014). Probabilistic Segmentation and Targeted Exploration of Objects in Cluttered Environments, IEEE Transactions on Robotics (TRo), 30, 5, pp.1198-1209.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  2. van Hoof, H.; Kroemer, O; Peters, J. (2013). Probabilistic Interactive Segmentation for Anthropomorphic Robots in Cluttered Environments , Proceedings of the International Conference on Humanoid Robots (HUMANOIDS).   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]
  3. Kroemer, O.; Lampert, C.H.; Peters, J. (2011). Learning Dynamic Tactile Sensing with Robust Vision-based Training, IEEE Transactions on Robotics (T-Ro), 27, 3, pp.545-557.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Interested in this robot system? Please contact Oleg Arenz!

Allegro hand

The Allegro hand has four fingers consisting of four joints each, giving the hand 16 degrees of freedom in total. This complexity enables the hand to accomplish dexterous manipulation tasks. It comes with a PD controller and is position controlled. The hand comes with sticky rubber sensorless fingertips that can grasp a variety of objects, but do not provide any sensory feedback. Thus, for in-hand manipulation tasks we equipped the hand with BioTac sensors. These are human inspired tactile fingertip sensors and can be seen on the picture here.
Find more information about the Allegro hand here.
Find more information about the BioTac tactile sensors here.

Some of our work with the Allegro:

  1. van Hoof, H.; Tanneberg, D.; Peters, J. (2017). Generalized Exploration in Policy Search, Machine Learning (MLJ), 106, 9-10, pp.1705-1724.   See Details [Details]   Download Article [PDF]   BibTeX Reference [BibTex]

Wessling Robotics Hand

The Wessling Robotic Hand, produced by Wessling Robotics, is composed of five robotic fingers. The hand is designed as to allow the fingers to be interchangeable. Each finger consists of three actuated degrees of freedom. One of these degrees of freedom is a coupled joint controlling both the distal and proximal joints. The hand then offers 15 actuated degrees of freedom and 20 joints in total. The fingertips of our Wessling Hand are equipped with BioTac SP sensors. These are a more recent version of the standard BioTac sensors produced by Syntouch. For our purposes, the hand is position controlled in joint space or controlled in task space with end-effectors placed at each fingertip.
For more information concerning the Wessling Robotic Hand please visit the Wessling Robotics website. Additional information on the BioTac SP tactile sensors can be found in the Syntouch website.

aDDa: Autonomous Driving Darmstadt 4 Students

IAS is also participating in the aDDa4Students https://www.tu-darmstadt.de/adda/adda project which is an interdisciplinary project, which aims to build an autonomous car at TU Darmstadt. Along with multiple other departments from the TU Darmstadt IAS we have access to a Mercedes S-Klasse including LIDAR, camera, radar and GPS sensors. The aDDa project hereby provides students the chance to implement, evaluate and enhance state of the art algorithms under realworld conditions.

If you are interested in working with the aDDa Car please contact Dorothea Koert!


zum Seitenanfang