This paper reports the results and post-challenge analyses of ChaLearn's AutoDL challenge series, which helped sorting out a profusion of AutoML solutions for Deep Learning (DL) that had been introduced in a variety of settings, but lacked fair comparisons. All input data modalities (time series, images, videos, text, tabular) were formatted as tensors and all tasks were multi-label classification problems. Code submissions were executed on hidden tasks, with limited time and computational resources, pushing solutions that get results quickly. In this setting, DL methods dominated, though popular Neural Architecture Search (NAS) was impractical. Solutions relied on fine-tuned pre-trained networks, with architectures matching data modality. Post-challenge tests did not reveal improvements beyond the imposed time limit. While no component is particularly original or novel, a high level modular organization emerged featuring a "meta-learner", "data ingestor", "model selector", "model/learner", and "evaluator". This modularity enabled ablation studies, which revealed the importance of (off-platform) meta-learning, ensembling, and efficient data management. Experiments on heterogeneous module combinations further confirm the (local) optimality of the winning solutions. Our challenge legacy includes an ever-lasting benchmark (http://autodl.chalearn.org), the open-sourced code of the winners, and a free "AutoDL self-service."

The planning and execution of real-world robotic tasks largely depends on the ability to generate feasible motions online in response to changing environment conditions or goals. A spline deformation method is able to modify a given trajectory so that it matches the new boundary conditions, e.g., on positions, velocities, accelerations, etc. At the same time, the deformed motion preserves velocity, acceleration, jerk, or higher derivatives of motion profile of the precalculated trajectory. The deformed motion possessing such properties can be expressed by translation of original trajectory and spline interpolation. This spline decomposition considerably reduces the computational complexity and allows real-time execution. Formal feasibility guarantees are provided for the deformed trajectory and for the resulting torques. These guarantees are based on the special properties of Bernstein polynomials used for the deformation and on the structure of the chosen computed-torque control scheme. The approach is experimentally evaluated in a number of planar volleyball experiments using 3 degree-of-freedom robots and human participants.

Ongoing technological advances in the areas of computation, sensing, and mechatronics enable robotic-based systems to interact with humans in the real world. To succeed against a human in a competitive scenario, a robot must anticipate the human behavior and include it in its own planning framework. Then it can predict the next human move and counter it accordingly, thus not only achieving overall better performance but also systematically exploiting the opponent's weak spots. Pool is used as a representative scenario to derive a model-based planning and control framework where not only the physics of the environment but also a model of the opponent is considered. By representing the game of pool as a Markov decision process and incorporating a model of the human decision-making based on studies, an optimized policy is derived. This enables the robot to include the opponent's typical game style into its tactical considerations when planning a stroke. The results are validated in simulations and real-life experiments with an anthropomorphic robot playing pool against a human.

This paper discusses an online dynamic motion generation scheme for nonprehensile object manipulation by using a set of predefined motions and a trajectory deformation algorithm capable of incorporating positional and velocity boundary constraints. By creating optimal trajectories offline and deforming them online, computational complexity during execution is reduced considerably. As tight convex hulls of the deformed trajectories can be found, possible obstacles or workspace boundaries can be circumnavigated precisely without collision. The approach is verified through experiments on an inclined planar air-table for volleyball scenario using two 3-DoF robots.

This paper presents a novel robot pose measure for human movement imitation based entirely on the Euclidean distance information between any two links of a robot and any link and object in the robot's environment in a Cartesian task space. A Hidden Markov Model is used to encode the spatio-temporal information of multiple demonstrations. In combination with Gaussian Mixture Regression for extracting the important task properties, feasible full-body motion adaption can be achieved. The method is suited for use with a humanoid robot by considering additional constraints like balance control and collision avoidance. In order to tackle modeling errors occurring due to the human movement demonstration and the robotic reproduction, a manipulability based weighting scheme is proposed. Complexity reduction of the otherwise redundant pose measure is performed based upon a mechanical analogy of an interconnected spring system. Experiments are conducted using a HRP-4 robot and display the applicability of the presented methods for robotic full-body motion imitation tasks.

This paper presents an incremental sampling-based approach for trajectory imitation in cluttered environments using the RRT* algorithm. Inspired by the discrete Laplace-Beltrami operator the underlying distance metric is based upon the difference from a reference trajectory through a quadratic distance term incorporating velocity and acceleration deviations along the trajectory. Mathematically-backed approximations in combination with a task-space bias make it possible to use standard nearest neighbor methods in task space when expanding the RRT*-tree. It is shown that metric-consistent biases considerably increase the convergence speed. The proposed approach is validated in simulations in a 2D environment and in experiments using a HRP-4 humanoid robot.

Adjusting to new situations by changing the shape of a prerecorded trajectory is an important aspect for robot manipulation in a constrained environment. For being recognized as a distinctive trajectory, the goal of any trajectory modification is to keep local and global properties as similar as possible compared to the reference trajectory. This paper presents a framework that can alter the shape of a trajectory by defining the position of a set of sampling points while maintaining local properties in a least-squares manner. The method consists of a three-staged approach first modifying the global shape of the trajectory and subsequently taking local features into account. Inspired by mesh processing used for 3D surface editing, differential coordinates based on the discretized Laplacian operator are used for measuring and maintaining local trajectory properties when deforming the trajectory. Last, a post-processing step based on a relaxed “as-rigid-as-possible” principle allows local deformations and length modifications of the trajectory for a better tradeoff between preserving local and global properties. Experiments verifying the applicability of the proposed algorithm are conducted using a 7-DoF anthropomorphic arm following a previously recorded and modified trajectory.

This paper presents a novel descriptor for trajectory classification in n dimensions, which is invariant with respect to scaling and rigid transformation. Using a hierarchical approach, the descriptor is able to capture both local and global features of the trajectory. The algorithm iteratively splits up every trajectory into smaller trajectory segments resulting in a binary tree. Inspired by the Frenet-Serret formulas, a projection onto a lower dimensional subspace is performed for every trajectory segment, providing a characteristic description of every trajectory. The subspace projection acts as a pseudo-curvature measure in every dimension. Successful applicability is shown through classification experiments in three and six dimensions using an RGB-D camera. For comparison with other algorithms, the Australian Sign Language dataset is also used for classification, showing a superior classification rate.

In comparison to industrial settings with structured environments, the operation of autonomous robots in unstructured and uncertain environments is more challenging. This video presents a generic control and system architecture ARCADE, applicable for real-time robot control in complex task situations. Several methods to cope with uncertainties are demonstrated with the example task of changing tires on a car. Approaches of object detection (applied to car, tires, and humans), robust real-time control of robot arms under perception uncertainty, and human-friendly haptic interaction are detailed. The video shows two robots jointly performing the task of mounting a mock-up tire to a real car using the proposed methods, realizing robust performance in an uncertain environment.

Aiming at the application in physical human-robot interaction, this paper presents a novel adaptive admittance control scheme for robotic manipulators. Special emphasis is drawn on the avoidance of oscillatory behavior in the presence of closed kinematic chains while keeping the rendered impedance low. The approach uses an online fast Fourier transform of the measured manipulator endeffector forces in order to detect oscillations and to adapt the admittance parameters dynamically. As a novel method towards human-centered control design the adaptation strategy is determined in a user study evaluated with a machine-learning algorithm. Experiments conducted with ten human participants show superiority over the non-adaptive admittance control scheme.

This video presents a robot capable of playing pool on a normal sized pool table using two arms. For successfully completing this task several issues need to be addressed, including the perception of relevant environment information, planning of actions and finally an efficient execution. The video outlines how the robot accurately locates the pool table, the balls on the table and the cue and subsequently plans the next shot. In order to improve the stroke speed, an optimization algorithm for the arm configuration is described. Finally, it is shown how all these modules are integrated to achieve a working two-handed robotic pool play.