Marc Bestmann

German Aerospace Center (DLR) | DLR · Institute of Maintenance Repair and Overhaul

This paper investigates the influence of reference motion quality and other design choices on the performance of deep reinforcement learning for bipedal walking with Proximate Policy Optimization (PPO). We use parametrized Cartesian quintic splines to generate reference actions for an omnidirectional walk policy. By using parameter sets with differ...

This paper presents a open-source omnidirectional walk controller that provides bipedal walking for non-parallel robots through parameter optimization. The approach relies on pattern generation with quintic splines in Cartesian space. Additionally, baselines of achieved walk velocities in simulation for all robots of the Humanoid Virtual Season, as...

This thesis investigates the learning of motion skills for humanoid robots. As ground- work, a humanoid robot with integrated fall management was developed as an experi- mental platform. Then, two different approaches for creating motion skills were investi- gated. First, one that is based on Cartesian quintic splines with optimized parameters. Sec...

This paper presents a open-source omnidirectional walk controller that provides bipedal walking for non-parallel robots through parameter optimization. The approach relies on pattern generation with quintic splines in Cartesian space. Additionally, baselines of achieved walk velocities in simulation for all robots of the Humanoid Virtual Season, as...

We present the Dynamic Stack Decider (DSD), a lightweight open-source control architecture. It combines different well-known approaches and is inspired by behavior trees as well as hierarchical state machines . The DSD allows to design and structure complex behavior of robots as well as software agents while providing easy maintainability. Challeng...

We present a dataset specifically designed to be used as a benchmark to compare vision systems in the RoboCup Humanoid Soccer domain. The dataset is composed of a collection of images taken in various real-world locations as well as a collection of simulated images. It enables comparing vision approaches with a meaningful and expressive metric. The...

Fast and accurate visual perception utilizing a robot's limited hardware resources is necessary for many mobile robot applications. We are presenting YOEO, a novel hybrid CNN which unifies previous object detection and semantic segmentation approaches using one shared encoder backbone to increase performance and accuracy. We show that it outperform...

Reliability and robustness to external influences are important characteristics to using humanoid robots outside of laboratory conditions This paper proposes a closed-loop system to recover from falling over and self-righten using quintic spline interpolation. By using PID controllers to correct divergence in the IMU readings, faster and more relia...

We present our open humanoid robot platform Wolfgang. The described hardware focuses on four aspects. Firstly, the robustness against falls is improved by integrating 3D printed elastic elements. Additionally, a high control loopfrequency is achieved by using new custom control electronics. Furthermore, a torsion spring is applied to reduce the tor...

We present a dataset specifically designed to be used as a benchmark to compare vision systems in the RoboCup Humanoid Soccer domain. The dataset is composed of a collection of images taken in various real-world locations as well as a collection of simulated images. It enables comparing vision approaches with a meaningful and expressive metric. The...

Humanoid robots promise a better integration of robots into our everyday life, but they come with additional challenges when compared with other mobile robots. We present a novel approach to simplify their usage by handling these challenges with the Humanoid Control Module and show its utility in the RoboCup Soccer Humanoid League. The paper also d...

We present the Dynamic Stack Decider (DSD), a lightweight open source control architecture. It combines different well-known approaches and is inspired by behavior trees as well as hierarchical state machines. The DSD allows to design and structure complex behavior of robots as well as software agents while providing easy maintainability. Challenge...

This extended abstract describes the current research of the Hamburg Bit-Bots Humanoid KidSize RoboCup team, lessons learned from last years competition and improvements planned for 2020. In the RoboCup 2019 competition, we had recently installed new cameras on the robots with a new 3D printed head. These heads sometimes broke when the robot fell....

We are proposing an Open Source ROS vision pipeline for the RoboCup Soccer context. It is written in Python and offers sufficient precision while running with an adequate frame rate on the hardware of kid-sized humanoid robots to allow a fluent course of the game. Fully Convolutional Neural Networks (FCNNs) are used to detect balls while convention...

High-frequency control loops are necessary to improve agility and reactiveness of robots. One of the common limiting bottlenecks is the communication with the hardware, i.e., reading of sensors values and writing of actuator commands. In this paper, we investigate the performance of devices using the widespread Robotis Dynamixel protocol via an RS-...

Convolutional Neural Networks (CNNs) have shown promising results for various computer vision tasks. Despite their success, localizing the ball in real-world RoboCup Soccer scenes is still challenging. Especially considering real-time requirements and the limited computing power of humanoid robots. Another important reason is the lack of training a...

The need for labeled training data for object recognition in RoboCup increased due to the spread of deep learning approaches. Creating large sets of training images from different environments and annotating the recorded objects is difficult for a single RoboCup team.

High-frequency control loops are necessary to improve agility and reactiveness of robots. One of the common limiting bottlenecks is the communication with the hardware, i.e., reading of sensors values and writing of actuator commands. In this paper, we investigate the performance of devices using the widespread Robotis Dynamixel protocol via an RS-...

We are proposing an Open Source ROS vision pipeline for the RoboCup Soccer context. It is written in Python and offers sufficient precision while running with an adequate frame rate on the hardware of kid-sized humanoid robots to allow a fluent course of the game. Fully Convolutional Neural Networks (FCNNs) are used to detect balls while convention...

This paper presents an approach for using an image-based heat map as measurement input of a particle filter. Pixels of the heat map are transformed into Cartesian space relative to the robot and regarded as single measurements. The approach uses a novel observation model to weight the particles accordingly to the heat map pixels. While this paper f...

This team description paper presents the developments made by the joint team Hamburg Bit-Bots & WF-Wolves. We present new software approaches we programmed and evaluated, like the Dynamic Stack Decider (DSD), our advances in image processing and our improvements to the walking engine. Additionally the newly developed foot pressure sensors as well a...

We present the Active Self Deciding Stack (ASDS), a lightweight behavior framework. It combines different well-known approaches and is inspired by behavior trees and hierarchical state machines. The approach proved its benefits over the last years in the RoboCup Humanoid Kid-Size League but can be used for compact high-level control of robot behavi...

The need for labeled training data for object recognition in RoboCup increased due to the spread of deep learning approaches. Creating large sets of training images from different environments and annotating the recorded objects is difficult for a single RoboCup team. This paper presents our tool ImageTagger which facilitates creating and sharing s...

The increase in software complexity and hardwarecosts forces robot soccer teams to collaborate. Instead of focusingon a specific architecture, we propose a set of common ROSmessages as interface definitions to encourage software exchange.Furthermore, we implemented fundamental utility packages forcollaborative play in mixed teams.

Sharing software modules between teams in the RoboCup Humanoid League is difficult since all teams use different frameworks. This leads to reimplementation of software which slows the research process. A common framework for the league would resolve this. Therefore, this thesis proposes a ROS-based architecture which is defined by a set of ROS mess...

In this paper a new robot is presented which was designed especially for RoboCup soccer. It is an approach to evolve from the standard Darwin based skeleton towards a robot with more human motion capabilities. Many new features were added to the robot to adapt it for the special requirements of RoboCup Soccer. Therefore, the interaction possibiliti...