Torsten Kröger

Yahav Avigal, Lars Berscheid, Tamim Asfour et al.

Folding garments reliably and efficiently is a long standing challenge in robotic manipulation due to the complex dynamics and high dimensional configuration space of garments. An intuitive approach is to initially manipulate the garment to a canonical smooth configuration before folding. In this work, we develop SpeedFolding, a reliable and efficient bimanual system, which given user-defined instructions as folding lines, manipulates an initially crumpled garment to (1) a smoothed and (2) a folded configuration. Our primary contribution is a novel neural network architecture that is able to predict pairs of gripper poses to parameterize a diverse set of bimanual action primitives. After learning from 4300 human-annotated and self-supervised actions, the robot is able to fold garments from a random initial configuration in under 120s on average with a success rate of 93%. Real-world experiments show that the system is able to generalize to unseen garments of different color, shape, and stiffness. While prior work achieved 3-6 Folds Per Hour (FPH), SpeedFolding achieves 30-40 FPH.

Jonas C. Kiemel, Torsten Kröger

In this paper, we present a learning-based approach that allows a robot to quickly follow a reference path defined in joint space without exceeding limits on the position, velocity, acceleration and jerk of each robot joint. Contrary to offline methods for time-optimal path parameterization, the reference path can be changed during motion execution. In addition, our approach can utilize sensory feedback, for instance, to follow a reference path with a bipedal robot without losing balance. With our method, the robot is controlled by a neural network that is trained via reinforcement learning using data generated by a physics simulator. From a mathematical perspective, the problem of tracking a reference path in a time-optimized manner is formalized as a Markov decision process. Each state includes a fixed number of waypoints specifying the next part of the reference path. The action space is designed in such a way that all resulting motions comply with the specified kinematic joint limits. The reward function finally reflects the trade-off between the execution time, the deviation from the desired reference path and optional additional objectives like balancing. We evaluate our approach with and without additional objectives and show that time-optimized path tracking can be successfully learned for both industrial and humanoid robots. In addition, we demonstrate that networks trained in simulation can be successfully transferred to a real robot.

Lars Berscheid, Torsten Kröger

We present Ruckig, an algorithm for Online Trajectory Generation (OTG) respecting third-order constraints and complete kinematic target states. Given any initial state of a system with multiple Degrees of Freedom (DoFs), Ruckig calculates a time-optimal trajectory to an arbitrary target state defined by its position, velocity, and acceleration limited by velocity, acceleration, and jerk constraints. The proposed algorithm and implementation allows three contributions: (1) To the best of our knowledge, we derive the first time-optimal OTG algorithm for arbitrary, multi-dimensional target states, in particular including non-zero target acceleration. (2) This is the first open-source prototype of time-optimal OTG with limited jerk and complete time synchronization for multiple DoFs. (3) Ruckig allows for directional velocity and acceleration limits, enabling robots to better use their dynamical resources. We evaluate the robustness and real-time capability of the proposed algorithm on a test suite with over 1,000,000,000 random trajectories as well as in real-world applications.

Tom P. Huck, Christoph Ledermann, Torsten Kröger

We introduce a novel simulation-based approach to identify hazards that result from unexpected worker behavior in human-robot collaboration. Simulation-based safety testing must take into account the fact that human behavior is variable and that human error can occur. When only the expected worker behavior is simulated, critical hazards can remain undiscovered. On the other hand, simulating all possible worker behaviors is computationally infeasible. This raises the problem of how to find interesting (i.e., potentially hazardous) worker behaviors given a limited number of simulation runs. We frame this as a search problem in the space of possible worker behaviors. Because this search space can get quite complex, we introduce the following measures: (1) Search space restriction based on workflow-constraints, (2) prioritization of behaviors based on how far they deviate from the nominal behavior, and (3) the use of a risk metric to guide the search towards high-risk behaviors which are more likely to expose hazards. We demonstrate the approach in a collaborative workflow scenario that involves a human worker, a robot arm, and a mobile robot.

Lars Berscheid, Pascal Meißner, Torsten Kröger

Robot learning of real-world manipulation tasks remains challenging and time consuming, even though actions are often simplified by single-step manipulation primitives. In order to compensate the removed time dependency, we additionally learn an image-to-image transition model that is able to predict a next state including its uncertainty. We apply this approach to bin picking, the task of emptying a bin using grasping as well as pre-grasping manipulation as fast as possible. The transition model is trained with up to 42000 pairs of real-world images before and after a manipulation action. Our approach enables two important skills: First, for applications with flange-mounted cameras, picks per hours (PPH) can be increased by around 15% by skipping image measurements. Second, we use the model to plan action sequences ahead of time and optimize time-dependent rewards, e.g. to minimize the number of actions required to empty the bin. We evaluate both improvements with real-robot experiments and achieve over 700 PPH in the YCB Box and Blocks Test.

Lars Berscheid, Christian Friedrich, Torsten Kröger

Robot learning is often simplified to planar manipulation due to its data consumption. Then, a common approach is to use a fully-convolutional neural network to estimate the reward of grasp primitives. In this work, we extend this approach by parametrizing the two remaining, lateral Degrees of Freedom (DoFs) of the primitives. We apply this principle to the task of 6 DoF bin picking: We introduce a model-based controller to calculate angles that avoid collisions, maximize the grasp quality while keeping the uncertainty small. As the controller is integrated into the training, our hybrid approach is able to learn about and exploit the model-based controller. After real-world training of 27000 grasp attempts, the robot is able to grasp known objects with a success rate of over 92% in dense clutter. Grasp inference takes less than 50ms. In further real-world experiments, we evaluate grasp rates in a range of scenarios including its ability to generalize to unknown objects. We show that the system is able to avoid collisions, enabling grasps that would not be possible without primitive adaption.

Jonas C. Kiemel, Torsten Kröger

This paper presents an approach for learning online generation of collision-free and torque-limited robot trajectories. In order to generate future motions, a neural network is periodically invoked. Based on the current kinematic state of the robot and the network output, a trajectory for the current time interval can be calculated. The main idea of our paper is to execute the computed motion only if a collision-free and torque-limited way to continue the trajectory is known. In practice, the motion computed for the current time interval is extended by a braking trajectory and simulated using a physics engine. If the simulated trajectory complies with all safety constraints, the computed motion is carried out. Otherwise, the braking trajectory calculated in the previous time interval serves as an alternative safe behavior. Given a task-specific reward function, the neural network is trained using reinforcement learning. The design of the action space used for reinforcement learning ensures that all computed trajectories comply with kinematic joint limits. For our evaluation, simulated humanoid robots and industrial robots are trained to reach as many randomly placed target points as possible. We show that our method reliably prevents collisions with static obstacles and collisions between the robot arms, while generating motions that respect both torque limits and kinematic joint limits. Experiments with a real robot demonstrate that safe trajectories can be generated in real-time.

Tom P. Huck, Christoph Ledermann, Torsten Kröger

During the planning phase of industrial robot workplaces, hazard analyses are required so that potential hazards for human workers can be identified and appropriate safety measures can be implemented. Existing hazard analysis methods use human reasoning, checklists and/or abstract system models, which limit the level of detail. We propose a new approach that frames hazard analysis as a search problem in a dynamic simulation environment. Our goal is to identify workplace hazards by searching for simulation sequences that result in hazardous situations. We solve this search problem by placing virtual humans into workplace simulation models. These virtual humans act in an adversarial manner: They learn to provoke unsafe situations, and thereby uncover workplace hazards. Although this approach cannot replace a thorough hazard analysis, it can help uncover hazards that otherwise may have been overlooked, especially in early development stages. Thus, it helps to prevent costly re-designs at later development stages. For validation, we performed hazard analyses in six different example scenarios that reflect typical industrial robot workplaces.

Tom P. Huck, Christoph Ledermann, Torsten Kröger

Industrial human-robot collaborative systems must be validated thoroughly with regard to safety. The sooner potential hazards for workers can be exposed, the less costly is the implementation of necessary changes. Due to the complexity of robot systems, safety flaws often stay hidden, especially at early design stages, when a physical implementation is not yet available for testing. Simulation-based testing is a possible way to identify hazards in an early stage. However, creating simulation conditions in which hazards become observable can be difficult. Brute-force or Monte-Carlo-approaches are often not viable for hazard identification, due to large search spaces. This work addresses this problem by using a human model and an optimization algorithm to generate high-risk human behavior in simulation, thereby exposing potential hazards. A proof of concept is shown in an application example where the method is used to find hazards in an industrial robot cell.

Jonas C. Kiemel, Torsten Kröger

We present an approach to learn fast and dynamic robot motions without exceeding limits on the position $θ$, velocity $\dotθ$, acceleration $\ddotθ$ and jerk $\dddotθ$ of each robot joint. Movements are generated by mapping the predictions of a neural network to safely executable joint accelerations. The neural network is invoked periodically and trained via reinforcement learning. Our main contribution is an analytical procedure for calculating safe joint accelerations, which considers the prediction frequency $f_N$ of the neural network. As a result, the frequency $f_N$ can be freely chosen and treated as a hyperparameter. We show that our approach is preferable to penalizing constraint violations as it provides explicit guarantees and does not distort the desired optimization target. In addition, the influence of the selected prediction frequency on the learning performance and on the computing effort is highlighted by various experiments.

Lars Berscheid, Pascal Meißner, Torsten Kröger

Flexible pick-and-place is a fundamental yet challenging task within robotics, in particular due to the need of an object model for a simple target pose definition. In this work, the robot instead learns to pick-and-place objects using planar manipulation according to a single, demonstrated goal state. Our primary contribution lies within combining robot learning of primitives, commonly estimated by fully-convolutional neural networks, with one-shot imitation learning. Therefore, we define the place reward as a contrastive loss between real-world measurements and a task-specific noise distribution. Furthermore, we design our system to learn in a self-supervised manner, enabling real-world experiments with up to 25000 pick-and-place actions. Then, our robot is able to place trained objects with an average placement error of 2.7 (0.2) mm and 2.6 (0.8)°. As our approach does not require an object model, the robot is able to generalize to unknown objects while keeping a precision of 5.9 (1.1) mm and 4.1 (1.2)°. We further show a range of emerging behaviors: The robot naturally learns to select the correct object in the presence of multiple object types, precisely inserts objects within a peg game, picks screws out of dense clutter, and infers multiple pick-and-place actions from a single goal state.

Jonas C. Kiemel, Pascal Meißner, Torsten Kröger

We present TrueRMA, a data-efficient, model-free method to learn cost-optimized robot trajectories over a wide range of starting points and endpoints. The key idea is to calculate trajectory waypoints in Cartesian space by recursively predicting orthogonal adaptations relative to the midpoints of straight lines. We generate a differentiable path by adding circular blends around the waypoints, calculate the corresponding joint positions with an inverse kinematics solver and calculate a time-optimal parameterization considering velocity and acceleration limits. During training, the trajectory is executed in a physics simulator and costs are assigned according to a user-specified cost function which is not required to be differentiable. Given a starting point and an endpoint as input, a neural network is trained to predict midpoint adaptations that minimize the cost of the resulting trajectory via reinforcement learning. We successfully train a KUKA iiwa robot to keep a ball on a plate while moving between specified points and compare the performance of TrueRMA against two baselines. The results show that our method requires less training data to learn the task while generating shorter and faster trajectories.

Jonas C. Kiemel, Robin Weitemeyer, Pascal Meißner et al.

We present TrueÆdapt, a model-free method to learn online adaptations of robot trajectories based on their effects on the environment. Given sensory feedback and future waypoints of the original trajectory, a neural network is trained to predict joint accelerations at regular intervals. The adapted trajectory is generated by linear interpolation of the predicted accelerations, leading to continuously differentiable joint velocities and positions. Bounded jerks, accelerations and velocities are guaranteed by calculating the range of valid accelerations at each decision step and clipping the network's output accordingly. A deviation penalty during the training process causes the adapted trajectory to follow the original one. Smooth movements are encouraged by penalizing high accelerations and jerks. We evaluate our approach by training a simulated KUKA iiwa robot to balance a ball on a plate while moving and demonstrate that the balancing policy can be directly transferred to a real robot.

David Puljiz, Erik Stöhr, Katharina S. Riesterer et al.

Hand guidance emerged from the safety requirements for collaborative robots, namely possessing joint-torque sensors. Since then it has proven to be a powerful tool for easy trajectory programming, allowing lay-users to reprogram robots intuitively. Going beyond, a robot can learn tasks by user demonstrations through kinesthetic teaching, enabling robots to generalise tasks and further reducing the need for reprogramming. However, hand guidance is still mostly relegated to collaborative robots. Here we propose a method that doesn't require any sensors on the robot or in the robot cell, by using a Microsoft HoloLens augmented reality head mounted display. We reference the robot using a registration algorithm to match the robot model to the spatial mesh. The in-built hand tracking and localisation capabilities are then used to calculate the position of the hands relative to the robot. By decomposing the hand movements into orthogonal rotations and propagating it down through the kinematic chain, we achieve a generalised hand guidance without the need to build a dynamic model of the robot itself. We tested our approach on a commonly used industrial manipulator, the KUKA KR-5.

Lars Berscheid, Pascal Meißner, Torsten Kröger

Robotic grasping in cluttered environments is often infeasible due to obstacles preventing possible grasps. Then, pre-grasping manipulation like shifting or pushing an object becomes necessary. We developed an algorithm that can learn, in addition to grasping, to shift objects in such a way that their grasp probability increases. Our research contribution is threefold: First, we present an algorithm for learning the optimal pose of manipulation primitives like clamping or shifting. Second, we learn non-prehensible actions that explicitly increase the grasping probability. Making one skill (shifting) directly dependent on another (grasping) removes the need of sparse rewards, leading to more data-efficient learning. Third, we apply a real-world solution to the industrial task of bin picking, resulting in the ability to empty bins completely. The system is trained in a self-supervised manner with around 25000 grasp and 2500 shift actions. Our robot is able to grasp and file objects with 274 picks per hour. Furthermore, we demonstrate the system's ability to generalize to novel objects.

David Puljiz, Katharina S. Riesterer, Björn Hein et al.

Having a precise and robust transformation between the robot coordinate system and the AR-device coordinate system is paramount during human-robot interaction (HRI) based on augmented reality using Head mounted displays (HMD), both for intuitive information display and for the tracking of human motions. Most current solutions in this area rely either on the tracking of visual markers, e.g. QR codes, or on manual referencing, both of which provide unsatisfying results. Meanwhile a plethora of object detection and referencing methods exist in the wider robotic and machine vision communities. The precision of the referencing is likewise almost never measured. Here we would like to address this issue by firstly presenting an overview of currently used referencing methods between robots and HMDs. This is followed by a brief overview of object detection and referencing methods used in the field of robotics. Based on these methods we suggest three classes of referencing algorithms we intend to pursue - semi-automatic, on-shot; automatic, one-shot; and automatic continuous. We describe the general workflows of these three classes as well as describing our proposed algorithms in each of these classes. Finally we present the first experimental results of a semi-automatic referencing algorithm, tested on an industrial KUKA KR-5 manipulator.

Lars Berscheid, Thomas Rühr, Torsten Kröger

Given the task of learning robotic grasping solely based on a depth camera input and gripper force feedback, we derive a learning algorithm from an applied point of view to significantly reduce the amount of required training data. Major improvements in time and data efficiency are achieved by: Firstly, we exploit the geometric consistency between the undistorted depth images and the task space. Using a relative small, fully-convolutional neural network, we predict grasp and gripper parameters with great advantages in training as well as inference performance. Secondly, motivated by the small random grasp success rate of around 3%, the grasp space was explored in a systematic manner. The final system was learned with 23000 grasp attempts in around 60h, improving current solutions by an order of magnitude. For typical bin picking scenarios, we measured a grasp success rate of 96.6%. Further experiments showed that the system is able to generalize and transfer knowledge to novel objects and environments.

David Puljiz, Erik Stöhr, Katharina S. Riesterer et al.

Hand guidance of robots has proven to be a useful tool both for programming trajectories and in kinesthetic teaching. However hand guidance is usually relegated to robots possessing joint-torque sensors (JTS). Here we propose to extend hand guidance to robots lacking those sensors through the use of an Augmented Reality (AR) device, namely Microsoft's Hololens. Augmented reality devices have been envisioned as a helpful addition to ease both robot programming and increase situational awareness of humans working in close proximity to robots. We reference the robot by using a registration algorithm to match a robot model to the spatial mesh. The in-built hand tracking capabilities are then used to calculate the position of the hands relative to the robot. By decomposing the hand movements into orthogonal rotations we achieve a completely sensorless hand guidance without any need to build a dynamic model of the robot itself. We did the first tests our approach on a commonly used industrial manipulator, the KUKA KR-5.