RSS 2019 论文列表

This paper develops an exploration framework that leverages Gaussian mixture models (GMMs) for high-fidelity perceptual modeling and exploits the compactness of the distributions for information sharing in communications-constrained applications. State-of-the-art, high-resolution perceptual modeling techniques do not always consider the implications of transferring the model across limited bandwidth communications channels, which is critical for real time information sharing. To bridge this gap in the state of the art, this paper presents a system that compactly represents sensor observations as GMMs and maintains a local occupancy grid map for a sampling-based motion planner that maximizes an information-theoretic objective function. The method is extensively evaluated in long duration simulations on an embedded PC and deployed to an aerial robot equipped with a 3D LiDAR. The result is significant memory efficiency as compared to state-of-the-art techniques.

We propose a solution to a time-varying variant of Markov Decision Processes which can be used to address the decision-theoretic planning problems for autonomous systems operating in unstructured outdoor environments. We explore the time variability property of the planning stochasticity and investigate the state reachability in order to design an efficient method that can well trade-off the solution optimality and time complexity. The reachability space is constructed by analyzing the means and variances of future states' reaching time. We validate our algorithm through extensive simulations using ocean data and the results show that our method has a great performance in terms of both solution quality and computing time.

In motion planning problems for autonomous robots, such as self-driving cars, the robot must ensure that its planned path is not in close proximity to obstacles in the environment. However, the problem of evaluating the proximity is generally non-convex and serves as a significant computational bottleneck for motion planning algorithms. In this paper, we present methods for a general class of absolutely continuous parametric curves to compute: (i) the minimum separating distance, (ii) tolerance verification, and (iii) collision detection. Our methods efficiently compute bounds on obstacle proximity by bounding the curve in a convex region. This bound is based on an upper bound on the curve arc length that can be expressed in closed form for a useful class of parametric curves including curves with trigonometric or polynomial bases. We demonstrate the computational efficiency and accuracy of our approach through numerical simulations of several proximity problems.

Human-robot interaction benefits greatly from multimodal sensor inputs as they enable increased robustness and generalization accuracy. Despite this observation, few HRI methods are capable of efficiently performing inference for multimodal systems. In this work, we introduce a reformulation of Interaction Primitives which allows for learning from demonstration of interaction tasks, while also gracefully handling nonlinearities inherent to multimodal inference in such scenarios. We also empirically show that our method results in more accurate, more robust, and faster inference than standard Interaction Primitives and other common methods in challenging HRI scenarios.

As the adoption of robots becomes widespread across more industries and domains, those robots will be placed in new contexts where they will interact with people who do not understand how they work. The consequences of such a disparity can already be seen in how people assign anthropomorphic characteristics to those robots, despite what what the robot designers may have intended. In this paper, we seek to understand how people interpret a social robot's performance of an emotion, what we term 'affective display,' and the positive or negative valence of that affect. To this end, we tasked annotators with observing the Anki Cozmo robot perform its over 900 pre-scripted behaviors and labeling those behaviors with 16 possible affective display labels (e.g., interest, boredom, disgust, etc.). In our first experiment, we trained a neural network to predict annotated labels given multimodal information about the robot's movement, face, and audio. The results suggest that pairing affects to predict the valence between them is more informative, which we confirmed in a second experiment. Both experiments show that certain modalities are more useful for predicting displays of affect and valence. For our final experiment, we generated novel robot behaviors and tasked human raters with assigning scores to valence pairs instead of applying labels, then compared our model's predictions of valence between the affective pairs and compared the results to the human ratings. We conclude that robot designers and researchers cannot assume that people will perceive affect or valence as designed, and make several suggestions for directions of future work.

Tracking 6D poses of objects from videos provides rich information to a robot in performing different tasks such as manipulation and navigation. In this work, we formulate the 6D object pose tracking problem in the Rao-Blackwellizedparticle filtering framework, where the 3D rotation and the 3D translation of the object are decoupled in the estimation process.This factorization allows our approach, called PoseRBPF, to efficiently estimate an object's 3D translation along with the full distribution over the 3D rotation. This is achieved by discretizing the rotation space in a fine-grained manner, and training an auto-encoder network to construct a codebook of feature embeddings for the discretized rotations. As a result, PoseRBPF can track objects with arbitrary symmetries while still maintaining adequate posterior distributions. Our approach achieves state-of-the-artresults on two 6D pose estimation benchmarks

Often times, we specify tasks for a robot using temporal language that can also span different levels of abstraction. The example command "go to the kitchen before going to the second floor" contains spatial abstraction, given that "floor" consists of individual rooms that can also be referred to in isolation ("kitchen", for example). There is also a temporal ordering of events, defined by the word "before". Previous works have used Linear Temporal Logic (LTL) to interpret temporal language (such as "before"), and Abstract Markov Decision Processes (AMDPs) to interpret hierarchical abstractions (such as "kitchen" and "second floor"), separately. To handle both types of commands at once, we introduce the Abstract Product Markov Decision Process (AP-MDP), a novel approach capable of representing non-Markovian reward functions at different levels of abstractions. The AP-MDP framework translates LTL into its corresponding automata, creates a product Markov Decision Process (MDP) of the LTL specification and the environment MDP, and decomposes the problem into subproblems to enable efficient planning with abstractions. AP-MDP performs faster than a non-hierarchical method of solving LTL problems in over 95% of tasks, and this number only increases as the size of the environment domain increases. We also present a neural sequence-to-sequence model trained to translate language commands into LTL expression, and a new corpus of non-Markovian language commands spanning different levels of abstraction. We test our framework with the collected language commands on a drone, demonstrating that our approach enables a robot to efficiently solve temporal commands at different levels of abstraction.

In many environmental monitoring scenarios, the sampling robot needs to simultaneously explore the environment and exploit features of interest with limited time. We present an anytime multi-objective informative planning method called Pareto Monte Carlo Tree Search which allows the robot to handle potentially competing objectives such as exploration versus exploitation. The method produces optimized decision solutions for the robot based on its knowledge (estimation) of the environment state, leading to better adaptation to environmental dynamics. We provide algorithmic analysis on the critical tree node selection step and show that the number of times choosing sub-optimal nodes is logarithmically bounded and the search result converges to the optimal choices at a polynomial rate.

In this paper, we address motion efficiency in autonomous robot exploration with multi-legged walking robots that can traverse rough terrains at the cost of lower efficiency and greater body vibration. We propose a robotic system for online and incremental learning of the terrain traversal cost that is immediately utilized to reason about next navigational goals in building spatial model of the robot surrounding. The traversal cost experienced by the robot is characterized by incrementally constructed Gaussian Processes using Bayesian Committee Machine. During the exploration, the robot builds the spatial terrain model, marks untraversable areas, and leverages the Gaussian Process predictive variance to decide whether to improve the spatial model or decrease the uncertainty of the terrain traversal cost. The feasibility of the proposed approach has been experimentally verified in a fully autonomous deployment with the hexapod walking robot.

This paper investigates the influence of different joint space and orientation representations on the approximation of the forward kinematics. We consider all degrees of freedom in three dimensional space SE(3) and in the robot's joint space Q. In order to approximate the forward kinematics, different shallow artificial neural networks with ReLU (rectified linear unit) activation functions are designed. The amount of weights and bias' of each network are normalized. The results show that quaternion/vector-pairs outperform other SE(3) representations with respect to the approximation capabilities, which is demonstrated with two robot types; a Stanford Arm and a concentric tube continuum robot. For the latter, experimental measurements from a robot prototype are used as well. Regarding measured data, if quaternion/vector-pairs are used, the approximation error with respect to translation and to rotation is found to be seven times and three times more accurate, respectively. By utilizing a four-parameter orientation representation, the position tip error is less than 0.8% with respect to the robot length on measured data showing higher accuracy compared to the state-of-the-art-modeling (1.5%) for concentric tube continuum robots. Other three-parameter representations of SO(3) cannot achieve this, for instance any sets of Euler angles (in the best case 3.5% with respect to the robot length).

Recent work has explored the problem of autonomous navigation by imitating a teacher and learning an end-to-end policy, which directly predicts controls from raw images. However, these approaches tend to be sensitive to mistakes by the teacher and do not scale well to other environments or vehicles. To this end, we propose Observational Imitation Learning (OIL), a novel imitation learning variant that supports online training and automatic selection of optimal behavior by observing multiple imperfect teachers. We apply our proposed methodology to the challenging problems of autonomous driving and UAV racing. For both tasks, we utilize the Sim4CV simulator that enables the generation of large amounts of synthetic training data and also allows for online learning and evaluation. We train a perception network to predict waypoints from raw image data and use OIL to train another network to predict controls from these waypoints. Extensive experiments demonstrate that our trained network outperforms its teachers, conventional imitation learning (IL) and reinforcement learning (RL) baselines and even humans in simulation.

Today's robotic systems are increasingly turning to computationally expensive models such as deep neural networks (DNNs) for tasks like localization, perception, planning, and object detection. However, resource-constrained robots, like low-power drones, often have insufficient on-board compute resources or power reserves to scalably run the most accurate, state-of-the art neural network compute models. Cloud robotics allows mobile robots the benefit of offloading compute to centralized servers if they are uncertain locally or want to run more accurate, compute-intensive models. However, cloud robotics comes with a key, often understated cost: communicating with the cloud over congested wireless networks may result in latency or loss of data. In fact, sending high data-rate video or LIDAR from multiple robots over congested networks can lead to prohibitive delay for real-time applications, which we measure experimentally. In this paper, we formulate a novel Robot Offloading Problem - how and when should robots offload sensing tasks, especially if they are uncertain, to improve accuracy while minimizing the cost of cloud communication? We formulate offloading as a sequential decision making problem for robots, and propose a solution using deep reinforcement learning. In both simulations and hardware experiments using state-of-the art vision DNNs, our offloading strategy improves vision task performance by between 1.3-2.3x of benchmark offloading strategies, allowing robots the potential to significantly transcend their on-board sensing accuracy but with limited cost of cloud communication.

Small-scale robots have the potential to impact many areas of medicine and manufacturing including targeted drug delivery, telemetry and micromanipulation. This paper develops an algorithmic framework for regulating external magnetic fields to induce motion in millimeter-scale robots in a viscous liquid, to simulate the physics of swimming at the micrometer scale. Our approach for planning motions for these swimmers is based on tools from geometric mechanics that provide a novel means to design periodic changes in the physical shape of a robot that propels it in a desired direction. Using these tools, we are able to derive new motion primitives for generating locomotion in these swimmers. We use these primitives for optimizing swimming efficiency as a function of its internal magnetization and describe a principled approach to encode the "best" magnetization distributions in the swimmers. We validate this procedure experimentally and conclude by implementing these newly computed motion primitives on several magnetic swimmer prototypes that include two-link and three-link swimmers.

Controlling soft robots with precision is a challenge due in large part to the difficulty of constructing models that are amenable to model-based control design techniques. Koopman operator theory offers a way to construct explicit linear dynamical models of soft robots and to control them using established model-based linear control methods. This method is data-driven, yet unlike other data-driven models such as neural networks, it yields an explicit control-oriented linear model rather than just a ``black-box'' input-output mapping. This work describes this Koopman-based system identification method and its application to model predictive controller design. A model and MPC controller of a pneumatic soft robot arm is constructed via the method, and its performance is evaluated over several trajectory following tasks in the real-world. On all of the tasks, the Koopman-based MPC controller outperforms a benchmark MPC controller based on a linear state-space model of the same system.

Many fundamental challenges in robotics, based in manipulation or locomotion, require making and breaking contact with the environment. To represent the complexity of frictional contact events, impulsive impact models are especially popular, as they often lead to mathematically and computationally tractable approaches. However, when two or more impacts occur simultaneously, the precise sequencing of impact forces is generally unknown, leading to the potential for multiple possible outcomes. This simultaneity is far from pathological, and occurs in many common robotics applications. In this work, we propose an approach for resolving simultaneous frictional impacts, represented as a differential inclusion. Solutions to our model, an extension to multiple contacts of Routh's method, naturally capture the set of potential post-impact velocities. We prove that solutions to the presented model must terminate. This is, to the best of our knowledge, the first such guarantee for set-valued outcomes to simultaneous frictional impacts.

This paper presents a data-driven methodology for linear embedding of nonlinear systems. Utilizing structural knowledge of general nonlinear dynamics, the authors exploit the Koopman operator to develop a systematic, data-driven approach for constructing a linear representation in terms of higher order derivatives of the underlying nonlinear dynamics. With the linear representation, the nonlinear system is then controlled with an LQR feedback policy, the gains of which need to be calculated only once. As a result, the approach enables fast control synthesis. We demonstrate the efficacy of the approach with simulations and experimental results on the modeling and control of a tail-actuated robotic fish and show that the proposed policy is comparable to backstepping control. To the best of our knowledge, this is the first experimental validation of Koopman-based LQR control.

Lazy graph search algorithms are efficient at solving motion planning problems where edge evaluation is the computational bottleneck. These algorithms work by lazily computing the shortest potentially feasible path, evaluating edges along that path, and repeating until a feasible path is found. The order in which edges are selected is critical to minimizing the total number of edge evaluations: a good edge selector chooses edges that are not only likely to be invalid, but also eliminates future paths from consideration. We wish to learn such a selector by leveraging prior experience. We formulate this problem as a Markov Decision Process (MDP) on the state of the search problem. While solving this large MDP is generally intractable, we show that we can compute oracular selectors that can solve the MDP during training. With access to such oracles, we use imitation learning to find effective policies. If new search problems are sufficiently similar to problems solved during training, the learned policy will choose a good edge evaluation ordering and solve the motion planning problem quickly. We evaluate our algorithms on a wide range of 2D and 7D problems and show that the learned selector outperforms baseline commonly used heuristics.

Deep reinforcement learning (deep RL) holds the promise of automating the acquisition of complex controllers that can map sensory inputs directly to low-level actions. In the domain of robotic locomotion, deep RL could enable learning locomotion skills with minimal engineering and without an explicit model of the robot dynamics. Unfortunately, applying deep RL to real-world robotic tasks is exceptionally difficult, primarily due to poor sample complexity and sensitivity to hyperparameters. While hyperparameters can be easily tuned in simulated domains, tuning may be prohibitively expensive on physical systems, such as legged robots, that can be damaged through extensive trial-and-error learning. In this paper, we propose a sample-efficient deep RL algorithm based on maximum entropy RL that requires minimal per-task tuning and only a modest number of trials to learn neural network policies. We apply this method to learning walking gaits on a real-world Minitaur robot. Our method can acquire a stable gait from scratch directly in the real world in about two hours, without relying on any model or simulation, and the resulting policy is robust to moderate variations in the environment. We further show that our algorithm achieves state-of-the-art performance on simulated benchmarks with a single set of hyperparameters. Videos of training and the learned policy can be found on the project website.

This paper introduces the Logic-based Value Iteration Network (LVIN) framework, which combines imitation learning and logical automata to enable agents to learn complex behaviors from demonstrations. We address two problems with learning from expert knowledge: (1) how to generalize learned policies for a task to larger classes of tasks, and (2) how to account for erroneous demonstrations. Our LVIN model solves finite gridworld environments by instantiating a recurrent, convolutional neural network as a value iteration procedure over a learned Markov Decision Process (MDP) that factors into two MDPs: a small finite state automaton (FSA) corresponding to logical rules, and a larger MDP corresponding to motions in the environment. The parameters of LVIN (value function, reward map, FSA transitions, large MDP transitions) are approximately learned from expert trajectories. Since the model represents the learned rules as an FSA, the model is interpretable; since the FSA is integrated into planning, the behavior of the agent can be manipulated by modifying the FSA transitions. We demonstrate these abilities in several domains of interest, including a lunchbox-packing manipulation task and a driving domain.