ICRA 2015 Papers

On the one hand, mainly within the computer vision community, multi-resolution image labelling problems with pixel, super-pixel and object levels, have made great progress towards the modelling of holistic scene understanding. On the other hand, mainly within the robotics and graphics communities, multi-resolution 3D representations of the world have matured to be efficient and accurate. In this paper we bring together the two hands and move towards the new direction of unified recognition, reconstruction and representation. We tackle the problem by embedding an octree into a hierarchical robust PN Markov Random Field. This allows us to jointly infer the multi-resolution 3D volume along with the object-class labels, all within the constraints of an octree data-structure. The octree representation is chosen as this data-structure is efficient for further processing such as dynamic updates, data compression, and surface reconstruction. We perform experiments in inferring our semantic octree on the The kitti Vision Benchmark Suite in order to demonstrate its efficacy.

M-estimators are the de-facto standard method of robust estimation in robotics. They are easily incorporated into iterative non-linear least-squares estimation and provide seamless and effective handling of outliers in data. However, every M-estimator's robust loss function has one or more tuning parameters that control the influence of different data. The choice of M-estimator and the manual tuning of these parameters is always a source of uncertainty when applying the technique to new data or a new problem. In this paper we develop the concept of self-tuning M-estimators. We first make the connection between many common M-estimators and elliptical probability distributions. This connection shows that the choice of M-estimator is an assumption that the residuals belong to a well-defined elliptical distribution. We exploit this implication in two ways. First, we develop an algorithm for tuning the M-estimator parameters during iterative optimization. Second, we show how to choose the correct M-estimator for your data by examining the likelihood of the data given the model. We fully derive these algorithms and show their behavior on a representative example of visual simultaneous localization and mapping.

The implementation of collision detection mechanisms on-board of unmanned vehicles typically relies on digital platforms and shared communication networks. The discrete nature of these technologies does not only limit the frequency at which nearby obstacles and other vehicles are detected, but it also raises the demand for a better allocation of the vehicle's computational and communication resources. For instance, it is natural and convenient to adapt the frequency at which other vehicles are observed based on their distance and collision risk, a notion that relates to topics in minimum attention control and self-triggered control. In this paper, we report on a cooperative, decentralized self-triggered avoidance control law for an arbitrary large group of unmanned vehicles that reduces the attention and sampling frequency of other vehicles' position. The proposed control strategy guarantees the safe coordination of all vehicles at all times while reducing the sampling frequency of other vehicles' position based on their last available sample. We show that the sampling intervals are positive and lower bounded, which is critical for digital implementation. A simulation example is used to illustrate the performance of the avoidance control strategy.

The ability to learn about and efficiently use tools constitutes a desirable property for general purpose humanoid robots, as it allows them to extend their capabilities beyond the limitations of their own body. Yet, it is a topic that has only recently been tackled from the robotics community. Most of the studies published so far make use of tool representations that allow their models to generalize the knowledge among similar tools in a very limited way. Moreover, most studies assume that the tool is always grasped in its common or canonical grasp position, thus not considering the influence of the grasp configuration in the outcome of the actions performed with them. In the current paper we present a method that tackles both issues simultaneously by using an extended set of functional features and a novel representation of the effect of the tool use. Together, they implicitly account for the grasping configuration and allow the iCub to generalize among tools based on their geometry. Moreover, learning happens in a self-supervised manner: First, the robot autonomously discovers the affordance categories of the tools by clustering the effect of their usage. These categories are subsequently used as a teaching signal to associate visually obtained functional features to the expected tool's affordance. In the experiments, we show how this technique can be effectively used to select, given a tool, the best action to achieve a desired effect.

Actual control of most humanoid robots is based on the 3D linear inverted pendulum and assumes an horizontal displacement of the center of mass of the robot while obviously the center of mass in human walking is characterized by vertical oscillations. The objective of the paper is to show that these oscillations have a crucial role for the high level control of the walk. Based on a controlled length inverted pendulum model of the walker, it will be shown that vertical oscillations induce a self stabilization of the walk while this self stabilization is not observed in the case of a horizontal motion of the center of mass. The results are essentially based on the evolution of the angular momentum throughout the walk. The decrease of the angular momentum during the change of support is determinant to introduce a dependence between the path of the center of mass and the walking velocity. For a large set of walking characteristics (stride, velocity ...) a self synchronization of the motion in the sagittal and frontal planes appears that allows a low level control to produce stable cyclic gaits.

Self-assembling robots can be transported and deployed inexpensively and autonomously in remote and dangerous environments. In this paper, we introduce a novel self-assembling method with a planar pneumatic system. Inflation of pouches translate into shape changes, turning a sheet of composite material into a complex robotic structure. This new method enables a flat origami-based robotic structure to self-fold to desired angles with pressure control. It allows a static joint to become dynamic, self-actuate to reconfigure itself after initial folding. Finally, the folded robot can unfold itself at the end of a robotic application. We believe this new pneumatic approach provides an important toolkit to build more powerful and capable self-assembling robots.

The ability to selectively stiffen otherwise compliant soft actuators increases their versatility and dexterity. We investigate granular jamming and layer jamming as two possible methods to achieve stiffening with PneuFlex actuators, a type of soft continuum actuator. The paper details five designs of jamming compartments that can be attached to an actuator. We evaluate the stiffening of the five different prototypes, achieving an up to 8-fold increase in stiffness. The strength of the most effective prototype based on layer jamming is also validated in the context of pushing buttons, resulting in an 2.23-fold increase in pushing force.

In many real life scenarios, which span from domestic interactions to industrial manufacturing processes, the objects to be manipulated are non-rigid and deformable, hence, both the location of the object and its deformation have to be tracked. Different methodologies have been applied in literature, using different sensors and techniques for addressing this problem. The main contribution of this paper is to propose a Weightless Neural Network approach for non-rigid deformable object tracking. The proposed approach allows deploying an on-line training on the shape features of the object, to adapt in real-time to changes, and to partially cope with occlusions. Moreover, the use of parallel classifiers trained on the same set of images allows tracking the movements of the objects. In this work, we evaluate the filtering/segmentation performance that is a fundamental step for the correct operation of our approach, in the scenario of pizza making.

This paper proposes a new method for segmenting and classifying seamines on Synthetic Aperture Sonar (SAS) side scan images. The method uses an active contours approach and superellipse a-priori knowledge to segment the image in object, object-shadow and background areas. In contrast to other methods using superellipse constraints, the shape prior is incorporated directly into the segmentation process. This kind of segmentation has the advantage that afterwards the extracted superellipse parameters that describe the object and the object-shadow can directly be used as feature for a classification - this work is hence also of potential interest for general object recognition tasks in other application domains. Several different perspectives of implementing this idea into a suitable algorithm are introduced and compared with each other. Thus, for the evaluation of each method the extracted superellipse features are used for a support vector machine classification. An one against all confusion matrix is generated on a test data set. This result is compared to a related state of the art algorithm. It is shown that our new method is able to correctly classify 170 of 210 objects in a very challenging real world data set and that it yields significant better results than the state of the art comparison.

In this paper, we address the problem of motion planning for an autonomous grain cart. Contrary to prevailing practices in farming wherein one grain cart is allocated to each combine harvester, we envision a scenario in which one grain cart can serve multiple combine harvesters. This gives rise to challenging problems in scheduling and motion planning for the grain cart. In this work, we initially propose a scheduling scheme for a single grain cart to unload multiple combine harvesters without any interruption in the activity of the combines. The proposed algorithm can accommodate arbitrary number of combines. Based on the scheduling scheme, a planner is used to synthesize a feasible path for the grain cart. Finally, simulation results are presented to validate the feasibility of the proposed technique.

This paper explores the idea of reducing a robot's energy consumption while following a trajectory by turning off the main localisation subsystem and switching to a lower-powered, less accurate odometry source at appropriate times. This applies to scenarios where the robot is permitted to deviate from the original trajectory, which allows for energy savings. Sensor scheduling is formulated as a probabilistic belief planning problem. Two algorithms are presented which generate feasible perception schedules: the first is based upon a simple heuristic; the second leverages dynamic programming to obtain optimal plans. Both simulations and real-world experiments on a planetary rover prototype demonstrate over 50% savings in perception-related energy, which translates into a 12% reduction in total energy consumption.

As the number of rocket bodies and other debris in Earth's orbit increases, the need to capture and remove this space junk becomes essential to protect new satellites. A low cost solution may include gecko-inspired directional adhesives, which require almost no compressive preload to generate adhesion and are therefore suitable for surface grasping in space where objects are free floating. Current individual adhesive units with a pair of opposed pads achieve a limit of 13N normal to the surface. Instead of using a single large unit to generate high levels of adhesion, using multiple small gripper units is desirable to prevent single-point failures and to conform to higher curvatures. For this strategy to succeed, it is essential to distribute the overall force evenly, to minimize the overall preload normal to the surface, and to prevent local failures from propagating over the array. We present two load sharing mechanisms. The first uses nearly-constant force springs in parallel. The second uses a tendon and pulleys in series. Both allow a 4-unit gripper to maintain the same adhesive stress as a single unit. A normal adhesive load to compressive preload ratio of 100:1 is demonstrated. Zero gravity experiments and air bearing floor experiments demonstrate the gripper's functionality in a simulated space environment. Design considerations are discussed for further scaling, with the trade-offs among load sharing, suitability for different surfaces, and failure sensitivity.

Three-dimensional (3D) measurement is an important means for robots to acquire information about their environment. Structure from Motion is one of these 3D measurement methods. The 3D reconstruction of objects in the environment can be obtained from pictures captured with single camera in Structure from Motion. Furthermore, the camera motion can be obtained simultaneously. Because of its simplicity, Structure from Motion has been implemented in various ways. However there is an essential problem in that the scale of the measured objects cannot be computed by Structure from Motion. In order to compute the absolute scale, other information is required. However this is difficult for robots in an unknown situation. In this paper, we propose a method that can reconstruct the absolute scale of objects using refraction. Refraction changes the light ray path between the objects and the camera. This method is implemented using only a refractive plate and single camera. The results of simulations show the effectiveness of the proposed method in both air and other media (e.g., water).

Random sampling-based methods for motion planning of constrained robot manipulators have been widely studied in recent years. The main problem to deal with is the lack of an explicit parametrization of the non linear submanifold in the Configuration Space (CS) imposed by the constraints in the system. Most of the proposed planning methods use projections to generate valid configurations of the system slowing the planning process. Recently, new robot mechanism includes compliance either in the structure or in the controllers. In this kind of robot most of the times the planned trajectories are not executed exactly due to uncertainties and interactions with the environment. Indeed, controller references are generated such that the constraint is violated to indirectly generate forces during interactions. With the purpose of avoiding projections, in this paper we take advantage of the compliance of systems to relax the geometric constraints imposed by closed kinematic chains. The relaxed constraint is then used in a state-of-the-art suboptimal random sampling based technique to generate paths for constrained robot manipulators. As a consequence of relaxation, arising contact forces acting on the constraint change from configuration to configuration during the planned path. Those forces can be regulated using a proper controller that takes advantage of the geometric decoupling of the subspaces describing constrained rigid-body motions of the mechanism and the controllable forces.

The goal of saliency detection is to highlight objects in image data that stand out relative to their surrounding. Therefore, saliency detection aims to capture regions that are perceived as important. The most recent bottom-up approaches for saliency detection measure contrast based on visual features in 2D scenes, ignoring depth value. This work presents an effective method to measure saliency by mapping pixels into foreground and background regions in RGB-D images. Namely, we first segment an image into regions to evaluate the object uniqueness and consistency using graph-based segmentation. Then, we utilize the region color, depth, layout and boundary information to produce robust foreground and background saliency measures. Finally, we combine the two saliency maps based on Gaussian weights. As a result, our approach produces high-quality saliency maps, which may be used for further processing like object detection or recognition. Experimental results on two datasets compare our method with the state of the art and highlight its effectiveness.

We present a novel method based on saliency and segmentation to generate generic object candidates from RGB-D data. Our method uses saliency as a cue to roughly estimate the location and extent of the objects present in the scene. Salient regions are used to glue together the segments obtained from over-segmenting the scene by either color or depth segmentation algorithms, or by a combination of both. We suggest a late-fusion approach that first extracts segments from color and depth independently before fusing them to exploit that the data is complementary. Furthermore, we investigate several mechanisms for ranking the object candidates. We evaluate on one publicly available dataset and on one challenging sequence with a high degree of clutter. The results show that we are able to retrieve most objects in real-world indoor scenes and clearly outperform other state-of-the art methods.

A failsafe control strategy is presented for online safety certification of robot movements in a collaborative workspace with humans. This approach plans, predicts and uses formal guarantees on reachable sets of a robot arm and a human obstacle to verify the safety and feasibility of a trajectory in real time. The robots considered are serial link robots under Computed Torque schemes of control. We drastically reduce the computation time of our novel verification procedure through precomputation of non-linear terms and use of interval arithmetic, as well as representation of reachable sets by zonotopes, which scale easily to high dimensions and are easy to convert between joint space and Cartesian space. The approach is implemented in a simulation, to show that real time is computationally within reach.

This paper proposes a method called Segmentation-adaptive Pseudospectral collocation to address the problem of safe trajectory generation in missions with cooperating multiple aerial vehicles. Pseudospectral collocation can generate optimized collision-free trajectories, but for multiple aerial vehicles it cannot guarantee that the safety separation distance is maintained in the whole trajectories, since the constraints are only enforced in discrete points in the trajectory (collocation points). Hp-adaptive pseudospectral collocation increases iteratively the number of collocation points and the degree of the approximating polynomial, but this may lead to an exponential increase of the computational load. The proposed method solves the problem by selectively adding new collocation points where they are needed, only in the segments with conflicts in each iteration, thus effectively reducing the number of collocation points and the computation time with respect to other pseudospectral collocation formulations. The proposed method allows both changes of speed and changes of heading for each aerial vehicle to guarantee the safety distance between them. Its computational load and scalability are studied in randomly generated scenarios. Moreover, a comparison with other method is presented. Several experiments to test the validity of the approach have been also carried out in the multivehicle aerial testbed of the Center for Advanced Aerospace Technologies.

A prerequisite to safe robot motion is to avoid Inevitable Collision States (ICS). However, the characterization of the ICS set is a challenge. Several approximation methods have been proposed, most of which either are overly conservative or fail to provide proper motion safety guarantees. In order to to improve safety guarantees, we build upon Viability Theory and adapt an algorithm designed to approximate the Viability Kernel, a concept similar to ICS. Our algorithm is applied first to a challenging static environment scenario. It is then extended to handle dynamic environments. Although it is not possible in general to ensure safety forever, we manage nonetheless to achieve infinite motion safety in two special cases.