ICRA 2020 Papers

Hybrid legged-wheeled robots such as the CEN-TAURO, are capable of varying their footprint polygon to carry out various agile motions. This property can be advantageous for wheeled-only planning in cluttered spaces, which is our focus. In this paper, we present an improved algorithm that builds upon our previously introduced preliminary footprint varying A* planner, which was based on the rectangular symmetry of the foot support polygon. In particular, we introduce a Theta* based planner with trapezium-like search, which aims to further reduce the limitations imposed upon the wheeled-only navigation of the CENTAURO robot by the low-dimensional search space, maintaining the real-time computational efficiency. The method is tested on the simulated and real full-size CENTAURO robot in cluttered environments.

Rotating miniature magnetic swimmers are de-vices that could navigate within the bloodstream to access remote locations of the body and perform minimally invasive procedures. The rotational movement could be used, for example, to abrade a pulmonary embolus. Some regions, such as the heart, are challenging to navigate. Cardiac and respiratory motions of the heart combined with a fast and variable blood flow necessitate a highly agile swimmer. This swimmer should minimize contact with the walls of the blood vessels and the cardiac structures to mitigate the risk of complications. This paper presents experimental tests of a millimeter-scale magnetic helical swimmer navigating in a blood-mimicking solution and describes its turning capabilities. The step-out frequency and the position error were measured for different values of turn radius. The paper also introduces rapid movements that increase the swimmer's agility and demonstrates these experimentally on a complex 3D trajectory.

Optimal control holds great potential to improve a variety of robotic applications. The application of optimal control on-board limited platforms has been severely hindered by the large computational requirements of current state of the art implementations. In this work, we make use of a deep neural network to directly map the robot states to control actions. The network is trained offline to imitate the optimal control computed by a time consuming direct nonlinear method. A mixture of time optimality and power optimality is considered with a continuation parameter used to select the predominance of each objective. We apply our networks (termed G&CNets) to aggressive quadrotor control, first in simulation and then in the real world. We give insight into the factors that influence the `reality gap' between the quadrotor model used by the offline optimal control method and the real quadrotor. Furthermore, we explain how we set up the model and the control structure on-board of the real quadrotor to successfully close this gap and perform time-optimal maneuvers in the real world. Finally, G&CNet's performance is compared to state-of-the-art differential-flatness-based optimal control methods. We show, in the experiments, that G&CNets lead to significantly faster trajectory execution due to, in part, the less restrictive nature of the allowed state-to-input mappings.

This paper is about the closely-related problems of localization and aggregation for extremely simple robots, for which the only available action is to move in a given direction as far as the geometry of the environment allows. Such problems may arise, for example, in biomedical applications, wherein a large group of tiny robots moves in response to a shared external stimulus. Specifically, we extend the prior work on these kinds of problems presenting two algorithms for localization in environments with curved (rather than polygonal) boundaries and under low-friction models of interaction with the environment boundaries. We present both simulations and physical demonstrations to validate the approach.

Regrasping is one of the most common and important manipulation skills used in our daily life. However, aerial regrasping has not been seriously investigated yet, since most of the aerial manipulator lacks dexterous manipulation abilities except for the basic pick-and-place. In this paper, we focus on pivoting a long box, which is one of the most classical problems among regrasping researches, using a transformable multilink aerial robot. First, we improve our previous controller by compensating for the external wrench. Second, we optimize the joints configuration of our transformable multilink drone for stable grasping form under the constraints of thrust force and joints effort. Third, we sequentially optimize the grasping force in the pivoting process. The optimization goal is to generate continous grasping force whilst maximizing the friction force in case of the downwash, which would influence the grasped object and is difficult to model. Fourth, we develop the impedance controller in joint space and admittance controller in task space. As far as we know, it is the first research to achieve extrinsic contact-aware regrasping task on aerial robots.

Existing studies for environment interaction with an aerial robot have been focused on interaction with static surroundings. However, to fully explore the concept of an aerial manipulation, interaction with moving structures should also be considered. In this paper, a multirotor-based aerial manipulator opening a daily-life moving structure, a hinged door, is presented. In order to address the constrained motion of the structure and to avoid collisions during operation, model predictive control (MPC) is applied to the derived coupled system dynamics between the aerial manipulator and the door involving state constraints. By implementing a constrained version of differential dynamic programming (DDP), MPC can generate position setpoints to the disturbance observer (DOB)-based robust controller in real-time, which is validated by our experimental results.

Key challenges for the deployment of reinforcement learning (RL) agents in the real world are the discovery, representation and reuse of skills in the absence of a reward function. To this end, we propose a novel approach to learn a task-agnostic skill embedding space from unlabeled multi-view videos. Our method learns a general skill embedding independently from the task context by using an adversarial loss. We combine a metric learning loss, which utilizes temporal video coherence to learn a state representation, with an entropy-regularized adversarial skill-transfer loss. The metric learning loss learns a disentangled representation by attracting simultaneous viewpoints of the same observations and repelling visually similar frames from temporal neighbors. The adversarial skill-transfer loss enhances re-usability of learned skill embeddings over multiple task domains. We show that the learned embedding enables training of continuous control policies to solve novel tasks that require the interpolation of previously seen skills. Our extensive evaluation with both simulation and real world data demonstrates the effectiveness of our method in learning transferable skills from unlabeled interaction videos and composing them for new tasks. Code, pretrained models and dataset are available at http://robotskills.cs.uni-freiburg.de.

Deep reinforcement learning (RL) has enabled training action-selection policies, end-to-end, by learning a function which maps image pixels to action outputs. However, it's application to visuomotor robotic policy training has been limited because of the challenge of large-scale data collection when working with physical hardware. A suitable visuomotor policy should perform well not just for the task-setup it has been trained for, but also for all varieties of the task, including novel objects at different viewpoints surrounded by task-irrelevant objects. However, it is impractical for a robotic setup to sufficiently collect interactive samples in a RL framework to generalize well to novel aspects of a task. In this work, we demonstrate that by using adversarial training for domain transfer, it is possible to train visuomotor policies based on RL frameworks, and then transfer the acquired policy to other novel task domains. We propose to leverage the deep RL capabilities to learn complex visuomotor skills for uncomplicated task setups, and then exploit transfer learning to generalize to new task domains provided only still images of the task in the target domain. We evaluate our method on two real robotic tasks, picking and pouring, and compare it to a number of prior works, demonstrating its superiority.

When robots move autonomously for long-term, varied appearance such as the transition from day to night and seasonal variation brings challenges to visual place recognition. Defining an appearance condition (e.g. a season, a kind of weather) as a domain, we consider that the desired representation for place recognition (i) should be domain-unrelated so that images from different time can be matched regardless of varied appearance, (ii) should be learned in a self-supervised manner without the need of massive manually labeled data, and (iii) should be able to train among multiple domains in one model to keep limited model complexity. This paper sets to find domain-unrelated features across extremely changing appearance, which can be used as image descriptors to match between images collected at different conditions. We propose to use the adversarial network to disentangle domain-unrelated and domain-related features, which are named place and appearance features respectively. During training, only domain information is needed without requiring manually aligned image sequences. Experiments demonstrated that our method can disentangle place and appearance features in both toy case and images from the real world, and the place feature is qualified in place recognition tasks under different appearance conditions. The proposed network is also adaptable to multiple domains without increasing model capacity and shows favorable generalization.

In the autonomous driving area synthetic data is crucial for cover specific traffic scenarios which autonomous vehicle must handle. This data commonly introduces domain gap between synthetic and real domains. In this paper we deploy data augmentation to generate custom traffic scenarios with VRUs in order to improve pedestrian recognition. We provide a pipeline for augmentation of the Cityscapes dataset with virtual pedestrians. In order to improve augmentation realism of the pipeline we reveal a novel generative network architecture for adversarial learning of the data-set lighting conditions. We also evaluate our approach on the tasks of semantic and instance segmentation.

Path planning is an active area of research essential for many applications in robotics. Popular techniques include graph-based searches and sampling-based planners. These approaches are powerful but have limitations.This paper continues work to combine their strengths and mitigate their limitations using a unified planning paradigm. It does this by viewing the path planning problem as the two subproblems of search and approximation and using advanced graph-search techniques on a sampling-based approximation.This perspective leads to Advanced BIT*. ABIT* combines truncated anytime graph-based searches, such as ATD*, with anytime almost-surely asymptotically optimal sampling-based planners, such as RRT*. This allows it to quickly find initial solutions and then converge towards the optimum in an anytime manner. ABIT* outperforms existing single-query, sampling- based planners on the tested problems in ℝ4 and ℝ8, and was demonstrated on real-world problems with NASA/JPL-Caltech.

Informed sampling-based planning algorithms exploit problem knowledge for better search performance. This knowledge is often expressed as heuristic estimates of solution cost and used to order the search. The practical improvement of this informed search depends on the accuracy of the heuristic.Selecting an appropriate heuristic is difficult. Heuristics applicable to an entire problem domain are often simple to define and inexpensive to evaluate but may not be beneficial for a specific problem instance. Heuristics specific to a problem instance are often difficult to define or expensive to evaluate but can make the search itself trivial.This paper presents Adaptively Informed Trees (AIT*), an almost-surely asymptotically optimal sampling-based planner based on BIT*. AIT* adapts its search to each problem instance by using an asymmetric bidirectional search to simultaneously estimate and exploit a problem-specific heuristic. This allows it to quickly find initial solutions and converge towards the optimum. AIT* solves the tested problems as fast as RRT-Connect while also converging towards the optimum.

In this study, a visual shock absorber capable of adapting to free-fall objects with various weights and speeds is designed and realized. The key element is a magnetic gear to passively absorb shock in the moment of contact, which is difficult for traditional feedback control to deal with. The magnetic gear allows the seamless transfer of control from the non-contact state to the contact state. 1000 Hz high-speed visual object tracking is used for preparation with position and velocity control in the object non-contact state. In the moment of object contact, the high backdrivability of the magnetic gear response by hardware provides high responsiveness to external force. After the impact, the plastic deformation control of a parallel-expressed Maxwell model handles the contact state.

Studies on object rearrangement planning typically consider known objects. Some learning-based methods can predict the movement of an unknown object after single-step interaction, but require intermediate targets, which are generated manually, to achieve the rearrangement task. In this work, we propose a framework for unknown object rearrangement. Our system first models an object through a small-amount of identification actions and adjust the model parameters during task execution. We implement the proposed framework based on a low-cost tabletop robot (under 180 USD) to demonstrate the advantages of using a physics engine to assist action prediction. Experimental results reveal that after running our adaptive learning procedure, the robot can successfully arrange a novel object using an average of five discrete pushes on our tabletop environment and satisfy a precise 3.5 cm translation and 5° rotation criterion.

For multi-robot teams with heterogeneous capabilities, typical task allocation methods assign tasks to robots based on the suitability of the robots to perform certain tasks as well as the requirements of the task itself. However, in real-world deployments of robot teams, the suitability of a robot might be unknown prior to deployment, or might vary due to changing environmental conditions. This paper presents an adaptive task allocation and task execution framework which allows individual robots to prioritize among tasks while explicitly taking into account their efficacy at performing the tasks-the parameters of which might be unknown before deployment and/or might vary over time. Such a specialization parameter-encoding the effectiveness of a given robot towards a task-is updated on-the-fly, allowing our algorithm to reassign tasks among robots with the aim of executing them. The developed framework requires no explicit model of the changing environment or of the unknown robot capabilities-it only takes into account the progress made by the robots at completing the tasks. Simulations and experiments demonstrate the efficacy of the proposed approach during variations in environmental conditions and when robot capabilities are unknown before deployment.

Fixed-wing vertical take-off and landing (VTOL) aircraft pose a unique control challenge that stems from complex aerodynamic interactions between wings and rotors. Thus, accurate estimation of external forces is indispensable for achieving high performance flight. In this paper, we present a composite adaptive nonlinear tracking controller for a fixed- wing VTOL. The method employs online adaptation of linear force models, and generates accurate estimation for wing and rotor forces in real-time based on information from a three-dimensional airflow sensor. The controller is implemented on a custom-built fixed-wing VTOL, which shows improved velocity tracking and force prediction during the transition stage from hover to forward flight, compared to baseline flight controllers.

Collaborative robots and space manipulators contain significant joint flexibility. It complicates the control design, compromises the control bandwidth, and limits the tracking accuracy. The imprecise knowledge of the flexible joint dynamics compounds the challenge. In this paper, we present a new control architecture for controlling flexible-joint robots. Our approach uses a multi-layer neural network to approximate unknown dynamics needed for the feedforward control. The network may be viewed as a linear-in-parameter representation of the robot dynamics, with the nonlinear basis of the robot dynamics connected to the linear output layer. The output layer weights are updated based on the tracking error and the nonlinear basis. The internal weights of the nonlinear basis are updated by online backpropagation to further reduce the tracking error. To use time scale separation to reduce the coupling of the two steps - the update of the internal weights is at a lower rate compared to the update of the output layer weights. With the update of the output layer weights, our controller adapts quickly to the unknown dynamics change and disturbances (such as attaching a load). The update of the internal weights would continue to improve the converge of the nonlinear basis functions. We show the stability of the proposed scheme under the "outer loop" control, where the commanded joint position is considered as the control input. Simulation and physical experiments are conducted to demonstrate the performance of the proposed controller on a Baxter robot, which exhibits significant joint flexibility due to the series-elastic joint actuators.

Autonomous navigation through unknown and complex environments is a fundamental capability that is essential in almost all robotic applications. Optimal robot path planning is critical to enable efficient navigation. Path planning is a complex, compute and memory intensive task. Traditional methods employ either graph based search methods or sample based methods to implement path planning, which are sub-optimal and compute/memory-intensive. To this end, an Adaptive Directional Planner (ADP) algorithm is devised to achieve real-time, energy-efficient, memory-optimized, robust local path planning for enabling efficient autonomous navigation of mobile robots. The ADP algorithm ensures that the paths are optimal and kinematically-feasible. Further, the proposed algorithm is tested with different challenging scenarios verifying the functionality and robustness. The ADP algorithm implementation results demonstrate 40- 60X less number of nodes and 40 - 50X less execution time compared to the standard TP-RRT schemes, without compromising on accuracy. Finally, the algorithm has also been implemented as an accelerator for non-holonomic, multi-shape, small form factor mobile robots to provide a silicon solution with high performance and low memory footprint (28KB).

We propose Adaptive Curriculum Generation from Demonstrations (ACGD) for reinforcement learning in the presence of sparse rewards. Rather than designing shaped reward functions, ACGD adaptively sets the appropriate task difficulty for the learner by controlling where to sample from the demonstration trajectories and which set of simulation parameters to use. We show that training vision-based control policies in simulation while gradually increasing the difficulty of the task via ACGD improves the policy transfer to the real world. The degree of domain randomization is also gradually increased through the task difficulty. We demonstrate zero-shot transfer for two real-world manipulation tasks: pick-and-stow and block stacking. A video showing the results can be found at https://lmb.informatik.uni-freiburg.de/projects/curriculum/.