ICRA 2018 Papers

Combinatorial filters are discrete structures for modeling and reasoning about robotic systems. Such filters are of interest not only because of the potential for reduction of the computational power needed to execute the filter, but also for the insight they can sometimes provide into the information requirements of certain robotic tasks. It is known that the filter minimization problem -that is, for a given filter, to find a combinatorial filter with the minimal number of states among all filters with equivalent behavior-is NP-hard. Intuition might suggest that the well-known notion of bisimulation might be of direct use for this minimization problem. Indeed, the bisimilarity relation -the union of all bisimulation relations over the state space of the original filter-is an equivalence relation, and one might attempt to reduce a filter by merging states that are equivalent under this relation. This paper studies this relationship between bisimulation and combinatorial filter reduction. Specifically, we show that every filter minimization problem can be solved by computing a quotient of the input filter with some relation, but that for some filters, the bisimilarity relation is not the correct relation for this purpose. We also characterize the result of the bisimulation quotient operation as the solution to a different, stricter filter minimization problem, and identify several classes of filters for which a variant of bisimulation, called compatibility, can be used to minimize filters in polynomial time.

Recently, the centroidal momentum dynamics has received substantial attention to plan dynamically consistent motions for robots with arms and legs in multi-contact scenarios. However, it is also non convex which renders any optimization approach difficult and timing is usually kept fixed in most trajectory optimization techniques to not introduce additional non convexities to the problem. But this can limit the versatility of the algorithms. In our previous work, we proposed a convex relaxation of the problem that allowed to efficiently compute momentum trajectories and contact forces. However, our approach could not minimize a desired angular momentum objective which seriously limited its applicability. Noticing that the non-convexity introduced by the time variables is of similar nature as the centroidal dynamics one, we propose two convex relaxations to the problem based on trust regions and soft constraints. The resulting approaches can compute time-optimized dynamically consistent trajectories sufficiently fast to make the approach realtime capable. The performance of the algorithm is demonstrated in several multi-contact scenarios for a humanoid robot. In particular, we show that the proposed convex relaxation of the original problem finds solutions that are consistent with the original non-convex problem and illustrate how timing optimization allows to find motion plans that would be difficult to plan with fixed timing ††Implementation details and demos can be found in the source code available at https://git-amd.tuebingen.mpg.de/bponton/timeoptimization.

Extrinsic sensor calibration is an important task in robotics. There are various ways to perform the calibration task, but it often remains unclear which methods are better than the others. In this paper, we provide a systematic study about the calibration accuracy of three types of calibration methods, each represented by an abstract geometric model based on the sensor configuration and the calibration setup. We discuss the advantages and disadvantages of each model and perform a rigorous study on their noise sensitivity from a geometric perspective. As a result, we can reveal and quantify the relative calibration accuracies of the three models, thus answering the question of “which model is better and why?”. Beside our analytical analysis, we also provide numerical simulation experiments that validate our findings.

In this paper, we adapt the Continuous Collision Detection (CCD) method proposed in [1] to efficiently handle the case of spherical and two revolute joints, this kind of joints is very common in modern robotic systems. The new formulations provide more tight motion bounds, thus increase the success rate of checking collision-free paths. We also propose an extension to get the minimum distance to obstacles along a path, this information is primordial as it allows sampling-based motion planning techniques to sort collision-free paths according to their minimum clearance. We have integrated our implementation into a sampling-based motion planning technique and validated it through simulation and on the real Baxter research robot. The experiments revealed that the method not only does not miss any collision between the robot and the obstacles, but also the minimum distance extension provides the path with the maximum clearance at no additional computational cost.

Over the recent years, long-term place recognition has attracted an increasing attention to detect loops for largescale Simultaneous Localization and Mapping (SLAM) in loopy environments during long-term autonomy. Almost all existing methods are designed to work with traditional cameras with a limited field of view. Recent advances in omnidirectional sensors offer a robot an opportunity to perceive the entire surrounding environment. However, no work has existed thus far to research how omnidirectional sensors can help long-term place recognition, especially when multiple types of omnidirectional sensory data are available. In this paper, we propose a novel approach to integrate observations obtained from multiple sensors from different viewing angles in the omnidirectional observation in order to perform multi-directional place recognition in longterm autonomy. Our approach also answers two new questions when omnidirectional multisensory data is available for place recognition, including whether it is possible to recognize a place with long-term appearance variations when robots approach it from various directions, and whether observations from various viewing angles are the same informative. To evaluate our approach and hypothesis, we have collected the first large-scale dataset that consists of omnidirectional multisensory (intensity and depth) data collected in urban and suburban environments across a year. Experimental results have shown that our approach is able to achieve multi-directional long-term place recognition, and identifies the most discriminative viewing angles from the omnidirectional observation.

Visual place recognition is challenging, especially when only a few place exemplars are given. To mitigate the challenge, we consider place recognition method using omnidirectional cameras and propose a novel Omnidirectional Convolutional Neural Network (O-CNN) to handle severe camera pose variation. Given a visual input, the task of the O-CNN is not to retrieve the matched place exemplar, but to retrieve the closest place exemplar and estimate the relative distance between the input and the closest place. With the ability to estimate relative distance, a heuristic policy is proposed to navigate a robot to the retrieved closest place. Note that the network is designed to take advantage of the omnidirectional view by incorporating circular padding and rotation invariance. To train a powerful O-CNN, we build a virtual world for training on a large scale. We also propose a continuous lifted structured feature embedding loss to learn the concept of distance efficiently. Finally, our experimental results confirm that our method achieves state-of-the-art accuracy and speed with both the virtual world and real-world datasets.

Developing software for large scale off-road robot applications is challenging and tedious due to cost, logistics, and rigor of field testing. High-fidelity sensor-realistic simulation can speed up the development process for perception and state estimation algorithms. We focus on Lidar simulation for robots operating in off-road environments. Lidars are integral sensors for robots, and Lidar simulation for off-road environments is particularly challenging due to the way Lidar rays interact with natural terrain such as vegetation. A hybrid geometric terrain representation has been shown to model Lidar observations well [1]. However, previous work has only been able to simulate a single, fixed scene, and the entire scene had to be precisely surveyed. In this work, we add semantic information to the hybrid geometric model. This allows us to extract terrain primitives, such as trees and shrubs, from data logs. Our approach uses these primitives to compose arbitrary scenes for Lidar simulation. We evaluate our simulator on a real-world environment of interest, and show that primitives derived using our approach generalize to new scenes.

For many types of robots, avoiding obstacles is necessary to prevent damage to the robot and environment. As a result, obstacle avoidance has historically been an important problem in robot path planning and control. Soft robots represent a paradigm shift with respect to obstacle avoidance because their low mass and compliant bodies can make collisions with obstacles inherently safe. Here we consider the benefits of intentional obstacle collisions for soft robot navigation. We develop and experimentally verify a model of robot-obstacle interaction for a tip-extending soft robot. Building on the obstacle interaction model, we develop an algorithm to determine the path of a growing robot that takes into account obstacle collisions. We find that obstacle collisions can be beneficial for open-loop navigation of growing robots because the obstacles passively steer the robot, both reducing the uncertainty of the location of the robot and directing the robot to targets that do not lie on a straight path from the starting point. Our work shows that for a robot with predictable and safe interactions with obstacles, target locations in a cluttered, mapped environment can be reached reliably by simply setting the initial trajectory. This has implications for the control and design of robots with minimal active steering.

We propose an object-centric framework to label and predict human manipulation actions from observations of the object trajectories in 3D space. The goal is to lift the low-level sensor observation to a context specific human vocabulary. The low-level visual sensory input from a depth camera is processed into high-level descriptive action labels using a directed action graph representation. It is built based on the concepts of pre-computed Location Areas (LA), regions within a scene where an action typically occur, and Sector-Maps (SM), reference trajectories between the LAs. The framework consists of two stages, an offline teaching phase for graph generation, and an online action recognition phase that maps the current observations to the generated graph. This graph representation allows the framework to predict the most probable action from the observed motion in real-time and to adapt its structure whenever a new LA appears. Furthermore, the descriptive action labels enable not only a better exchange of information between a human and a robot but they allow also the robots to perform high-level reasoning. We present experimental results on real human manipulation actions using a system designed with this framework to show the performance of prediction and labeling that can be achieved.

Lameness in cattle is a health issue where gait is modified to minimise pain. Cattle are currently visually assessed for locomotion score, which provides the degree of lameness for individual animals. This subjective method is costly in terms of labour, and its level of accuracy and ability to detect small changes in locomotion that is critical for early detection of lameness and associated intervention. Current automatic lameness detection systems found in literature have not yet met the ultimate goal of widespread commercial adoption. We present a sensor configuration to record cattle kinematics towards automatic lameness detection. This configuration features four Time of Flight sensors to view cattle from above and from one side as they exit an automatic rotary milking dairy. Two dimensional near infrared images sampled from 223 cows passing through the system were used to train a Faster R-CNN to detect hooves (F1-score = 0.90) and carpal/tarsal joints (Fl-score = 0.85). The depth images were used to project these detected key points into Cartesian space where they were tracked to obtain individual trajectories per limb. The results show that kinematic gait features can be successfully obtained as a first and important step towards objective, accurate, automatic lameness detection.

This paper explores locally-optimal, efficient locomotion of a two-link planar robot balancing on a single, unactuated wheel. Because this model is essentially an acrobot mounted on a passive wheel, we name this model the acrollbot. By actuating an internal degree of freedom, the model can indirectly produce ground reaction forces yielding net accelerations and decelerations, to achieve locomotion. As with bipedal robot locomotion, this toy system is particularly challenging to control due to the need to balance continuously while controlling forward locomotion speed. However, unlike typical legged or rolling locomotion solutions, it is not immediately obvious how best to exploit actuation, internal reconfigurations, and motions to produce and control forward velocity along the ground, providing a useful benchmarking system for exploring optimization techniques. We use a direct collocation optimization framework to study this toy system, both to achieve a range of feasible locomotion solutions for nonintuitive dynamic robot models, and to investigate optimization of physical robot parameterizations, in the sense of improving locomotion efficiency. The framework and example presented throughout are designed with an aim toward bridging the gap between non-intuitive, data-driven optimization and model-based methods for design and control of underactuated and dynamically-stable locomotion.

The moving platform of a conventional parallel mechanism is typically connected to the ground link via serial kinematic chains or limbs. When interconnections are created among the limbs, they form coupling sub-chains that can provide additional constraints to the motion of the platform. The resulting complex mechanism offers the potential for improved performance and functionality of the overall mechanism. The basic building block of the parallel mechanism with coupling sub-chain is found to be a two-loop mechanism. This paper presents a screw theory based general method for synthesizing two-loop mechanism by adding an open chain on top of an existing closed chain for a specified combination of rotational and translational degrees of freedom of the end-effector. A comprehensive set of two-loop mechanisms have been presented for various combinations of rotational and translational degrees of freedom.

In this work, we propose a novel robot learning framework called Neural Task Programming (NTP), which bridges the idea of few-shot learning from demonstration and neural program induction. NTP takes as input a task specification (e.g., video demonstration of a task) and recursively decomposes it into finer sub-task specifications. These specifications are fed to a hierarchical neural program, where bottom-level programs are callable subroutines that interact with the environment. We validate our method in three robot manipulation tasks. NTP achieves strong generalization across sequential tasks that exhibit hierarchal and compositional structures. The experimental results show that NTP learns to generalize well towards unseen tasks with increasing lengths, variable topologies, and changing objectives.stanfordvl.github.io/ntp/.

Model-free deep reinforcement learning algorithms have been shown to be capable of learning a wide range of robotic skills, but typically require a very large number of samples to achieve good performance. Model-based algorithms, in principle, can provide for much more efficient learning, but have proven difficult to extend to expressive, high-capacity models such as deep neural networks. In this work, we demonstrate that neural network dynamics models can in fact be combined with model predictive control (MPC) to achieve excellent sample complexity in a model-based reinforcement learning algorithm, producing stable and plausible gaits that accomplish various complex locomotion tasks. We further propose using deep neural network dynamics models to initialize a model-free learner, in order to combine the sample efficiency of model-based approaches with the high task-specific performance of model-free methods. We empirically demonstrate on MuJoCo locomotion tasks that our pure model-based approach trained on just random action data can follow arbitrary trajectories with excellent sample efficiency, and that our hybrid algorithm can accelerate model-free learning on high-speed benchmark tasks, achieving sample efficiency gains of 3-5× on swimmer, cheetah, hopper, and ant agents. Videos can be found at https://sites.google.com/view/mbmf.

Swarm robotics refers to the implementation of swarm intelligence features like autonomy and self-organization to a collective of robots. This study focuses on the construction of a topological graph that represents both the magnitude and orientation of swarm interactions. Such structure is used for identifying global parameters like leadership and to derive a relationship between the distribution of interaction magnitudes and swarm parameters. Interaction magnitudes were derived from the trajectory distance between nearest neighbors and it was found that the distribution is able to differentiate between only a small subset of controllers, communication ranges and swarm sizes. Leader detection was based on the analysis of position vectors orientation in local neighborhoods. The method was successful at a 100% rate for 10 and 30 robots, while for 60 a minimum rate of 67% was obtained. Additionally, processing times never exceeded a simulation duration for swarms up to 30 robots, with the potential to parallelize for larger sizes.

Companion robots are more and more taking the role of caregivers for elderly people. Elderly people sometimes take the advice given by their family members or caregivers as a criticism. In this context, persuasive communication skills could be helpful. A social psychology theory called Regulatory Focus states that people have one of two inclinations when taking decisions: Promotion or Prevention Focus. Also, based on these inclinations, people can be influenced by the way the message is sent, including the speed of the speech and the amplitude of body gestures. In this paper, we analyze the influence of Regulatory Focus on a negotiation scenario, using 3 conditions: (1) a robot with a promotion behavior, (2) a robot with a prevention behavior, and (3) a robot with a neutral behavior. Our results support the results found in the psychology literature related to Regulatory Focus, suggesting that Promotion participants were more influenced by the robot showing a Promotion based behavior. Moreover, Prevention participants were more relaxed on the condition with the robot showing a Prevention based behavior, and accepted the biggest concession between the initial and final offer.

We consider the task of monitoring spatiotemporal phenomena in real-time by deploying limited sampling resources at locations of interest irrevocably and without knowledge of future observations. This task can be modeled as an instance of the classical secretary problem. Although this problem has been studied extensively in theoretical domains, existing algorithms require that data arrive in random order to provide performance guarantees. These algorithms will perform arbitrarily poorly on data streams such as those encountered in robotics and environmental monitoring domains, which tend to have spatiotemporal structure. We focus on the problem of selecting representative samples from phenomena with periodic structure and introduce a novel sample selection algorithm that recovers a near-optimal sample set according to any monotone submodular utility function. We evaluate our algorithm on a seven-year environmental dataset collected at the Martha's Vineyard Coastal Observatory and show that it selects phytoplankton sample locations that are nearly optimal in an information-theoretic sense for predicting phytoplankton concentrations in locations that were not directly sampled. The proposed periodic secretary algorithm can be used with theoretical performance guarantees in many real-time sensing and robotics applications for streaming, irrevocable sample selection from periodic data streams.

A key challenge in multi-robot and multi-agent systems is generating solutions that are robust to other self-interested or even adversarial parties who actively try to prevent the agents from achieving their goals. The practicality of existing works addressing this challenge is limited to only small-scale synchronous decision-making scenarios or a single agent planning its best response against a single adversary with fixed, procedurally characterized strategies. In contrast this paper considers a more realistic class of problems where a team of asynchronous agents with limited observation and communication capabilities need to compete against multiple strategic adversaries with changing strategies. This problem necessitates agents that can coordinate to detect changes in adversary strategies and plan the best response accordingly. Our approach first optimizes a set of stratagems that represent these best responses. These optimized stratagems are then integrated into a unified policy that can detect and respond when the adversaries change their strategies. The near-optimality of the proposed framework is established theoretically as well as demonstrated empirically in simulation and hardware.

Providing an efficient strategy to navigate safely through unsignaled intersections is a difficult task that requires determining the intent of other drivers. We explore the effectiveness of Deep Reinforcement Learning to handle intersection problems. Using recent advances in Deep RL, we are able to learn policies that surpass the performance of a commonly-used heuristic approach in several metrics including task completion time and goal success rate and have limited ability to generalize. We then explore a system's ability to learn active sensing behaviors to enable navigating safely in the case of occlusions. Our analysis, provides insight into the intersection handling problem, the solutions learned by the network point out several shortcomings of current rule-based methods, and the failures of our current deep reinforcement learning system point to future research directions.