In this paper, we present a comprehensive dataset comprising 37.9 hours of sensor data collected from humanoid robots, including 18.3 hours of walking and 2,519 recorded falls. This extensive dataset is a valuable resource for various robotics and machine learning applications. Leveraging this data, we propose RePro-TCN, a Temporal Convolutional Network (TCN) enhanced with two novel extensions: Relaxed Loss Formulation and Progressive Forecasting. Predicting falls is a critical capability in humanoid robotics for implementing countermeasures such as lunging or stopping the walk. Thanks to the new dataset, we train RePro-TCN and demonstrate its superiority over previous approaches under real-world conditions that were previously unattainable.

This work proposes a novel image dataset for multi-camera multi-object tracking and a framework that allows users to generate similar datasets. The dataset, called Synth- Yard-MCMOT-l, is the first of its kind to be generated in a virtual environment with the main focus on the tracking of trucks in yard logistics environments. The dataset consists of a total of 12,008 images generated by eight different cameras. The images contain 44,232 bounding boxes and segmentation masks and 52 individual tracks. Additionally, we provide a ninth camera, which is used to generate unified ground-truth information for the whole scene from an orthographic, top-down perspective comparable to a bird's eye or map-view. The purpose of this dataset is to provide yard management systems with relevant data, which can be employed when aiming to determine the exact position of a truck and specifically identifying which gateway or designated parking spot it is located in. The purpose of the repository is to enable researches to create unique use-case-specific multi-camera tracking datasets with the included dataset-generation pipeline. Initial benchmarks for single-camera tracking demonstrate a mean identification F1 score score of 0.96 and a mean multiple object tracking accuracy score of 0.94, laying the baseline for computing world coordinates via multi-camera multi-object tracking.

Multi-robot navigation and dynamic obstacle avoidance are challenging problems in robot learning. Recent advancements in Deep Reinforcement Learning (DRL) have demonstrated great potential in this area. Nonetheless, they often face challenges related to low sample efficiency. To overcome this challenge, some research proposes simulators that incorporate hardware acceleration. Although these simulators improve efficiency, they often lack the flexibility to generate diverse learning scenarios as often needed in multi-robot scenarios, where the different environments have varying numbers of agents.In this paper, we introduce MuRoSim, a multi-robot simulation for lidar-based navigation specifically designed for DRL applications. Due to its high level of abstraction, complete implementation in C++, and rigorous thread pool utilization, MuRoSim achieves high computational performance. We apply MuRoSim for training navigation policies for omnidirectional mobile robots equipped with lidar sensors using DRL. Finally, we conduct extensive Sim-to-Real experiments to confirm the realism of the simulator, by deploying the learned policy for dynamic navigation with up to six robots in numerous of real- world experiments.

This paper introduces evoBOT, a novel robot platform for research on highly dynamic locomotion and human-machine interaction. evoBOT is capable of performing complex tasks such as handovers or manipulation while moving at high speeds. We provide an overview of the robot's core features and the underlying design decisions on both the mechanical and the electronic level. Moreover, we propose a reinforcement learning (RL) based control approach for training highly dynamic motions that is evaluated on a first set of robotic tasks, including robust balancing and dynamic locomotion. Lastly, we conduct extensive benchmarking on the adopted sim-to-real methods and present an initial sim-to-real pipeline for first transfer of the trained policies to the real robot. To accelerate robotics research in this direction, the full simulation model of the robot is released as open-source.

How well simulation results can be transferred to the real world depends to a large extent on the sim-to-real gap that therefore should be as small as possible. In this work, this gap is reduced exemplarily for a robot with an omni-directional drive, which is challenging to simulate, utilizing machine learning methods. For this purpose, a motion capture system is first used to record a suitable data set of the robot's movements. Then, a model based on physical principles and observations is designed manually, which includes some unknown parameters that are learned based on the training dataset. Since the model is not differentiable, the evolutionary algorithms NSGA-II and -III are applied. Finally, by the presented approach, a significant reduction of the sim-to-real gap can be observed even at higher velocities above 2 m/s. The ablation study also shows that the elements beyond normal simulations, such as the engine simulation, and the machine learning approach, are essential for success.

Reactive stepping is an important utility to regain balance when bipedal walking motions are disturbed. This paper sheds light on the reasons for humanoid robots to fall down. It presents a method to calculate modifications of predefined foot placements with the objective to minimize deviations of the Zero Moment Point from a reference without interrupting the walk. The calculation is in closed-form, and is embedded into a well-evaluated preview controller with observer based on the 3D Linear Inverted Pendulum Mode (3D-LIPM). Experiments in simulation and on a physical robot prove the benefit of the proposed system.

This paper presents a vision system which is designed to be used by the research community in the Standard Platform League ( http://www.tzi.de/spl ) (SPL) and potentially in the Humanoid League ( http://www.robocuphumanoid.org ) (HL) of the RoboCup. It is real-time capable, robust towards lighting changes and designed to minimize calibration. We describe the structure of the processor along with major ideas behind object recognition. Moreover, we prove the benefit of the proposed system by assessing recorded image data on the robot hardware. The vision system has already been successfully employed with the NAO robot by Aldebaran Robotics ( http://www.aldebaran-robotics.com ) in prior RoboCup competitions as well as several minor events.

In this paper, a method for generating a natural human-like walking pattern of a humanoid robot is proposed. Inspired by biomechanical studies on human walking, we model the walking pattern of the robot with continuous and differ-entiable mathematical functions. The proposed walking pattern involves three-dimensional waist motion of the robot instead of restricting the waist motion to a two-dimensional planar surface. Considering the real environment includes not only unevenness but also some degree of inclination that may cause the walking to be unstable, this paper also presents a method of controlling the body posture of the humanoid robot. The posture controller utilizing ankle joints is based on sensory feedback and modifies the reference trajectories in real time in order to stabilize the robot. We have implemented the proposed walking pattern and control method on the humanoid robot NAO. Dynamic walking experiments have verified that the proposed scheme improves the walking stability of the robot.