RSS 2017 论文列表

As the spatial scale of robots decrease in multi-robot systems, collisions cease to be catastrophic events that need to be avoided at all costs. This implies that less conservative, coordinated control strategies can be employed, where collisions are not only tolerated, but can potentially be harnessed as an information source. In this paper, we follow this line of inquiry by employing collisions as a sensing modality that provides information about the robots' surroundings. We envision a collection of robots moving around with no sensors other than binary, tactile sensors that can determine if a collision occurred, and let the robots use this information to determine their locations. We apply a probabilistic localization technique based on mean-field approximations that allows each robot to maintain and update a probability distribution over all possible locations. Simulations and real multi-robot experiments illustrate the feasibility of the proposed approach, and demonstrate how collisions in multi-robot systems can indeed be employed as useful information sources.

Deep reinforcement learning has emerged as a promising and powerful technique for automatically acquiring control policies that can process raw sensory inputs, such as images, and perform complex behaviors. However, extending deep RL to real-world robotic tasks has proven challenging, particularly in safety-critical domains such as autonomous flight, where a trial-and-error learning process is often impractical. In this paper, we explore the following question: can we train vision-based navigation policies entirely in simulation, and then transfer them into the real world to achieve real-world flight without a single real training image? We propose a learning method that we call CAD2RL, which can be used to perform collision-free indoor flight in the real world while being trained entirely on 3D CAD models. Our method uses single RGB images from a monocular camera, without needing to explicitly reconstruct the 3D geometry of the environment or perform explicit motion planning. Our learned collision avoidance policy is represented by a deep convolutional neural network that directly processes raw monocular images and outputs velocity commands. This policy is trained entirely on simulated images, with a Monte Carlo policy evaluation algorithm that directly optimizes the network's ability to produce collision-free flight. By highly randomizing the rendering settings for our simulated training set, we show that we can train a policy that generalizes to the real world, without requiring the simulator to be particularly realistic or high-fidelity. We evaluate our method by flying a real quadrotor through indoor environments, and further evaluate the design choices in our simulator through a series of ablation studies on depth prediction. For supplementary video see: https://youtu.be/nXBWmzFrj5s

Human machine teaming has, for decades, been conceptualized as a function allocation (FA) or levels of autonomy (LOA) process: the human is suited for some tasks, while the machine is suitable for others, and as machines improve they take over duties previously assigned to humans. A wide variety of methods'including adaptive, adjustable, blended, supervisory and mixed initiative control, implemented discretely or continuously, as potential fields, as virtual fixture interfaces, or haptic interfaces'are derivatives of FA/LOA. We formalize FA/LOA (and all their derivatives) under a single mathematical formulation called classical shared control (CSC). Despite the widespread adoption of CSC, we prove that it fails to optimize human and robot agreement and intent if either the human or robot model displays 'intention ambiguity' (e.g., the human's intended goal is unclear or the robot finds multiple viable solutions). Practically, this suboptimality can manifest as unnecessary and unresolvable disagreement (an unnecessary deadlock). For instance, if the robot chooses to go left around an obstacle and the human chooses to go right, CSC only provides two solutions: freeze in place or collide with the obstacle (we provide a wide variety of failure examples in [52], https://arxiv.org/abs/1611.09490). We find that CSC suboptimality stems from arbitrating over model samples, rather than over models. Our key insight is thus to arbitrate over human and robot distributions; we prove this method optimizes human and robot agreement and intent and resolves deadlocking. Our key contribution is computationally efficient distribution arbitration: if the human and robot carry N^h_t and N^R_t 'intentions,' the joint (naively) has N^hTN^R_t intentions. In our approach, deadlock solutions have vanishingly small coefficients and only Nmin = arg min{N^h_t,N^R_t } non-zero coefficients remain: our joint has fewer modes than the individual agent models. We call our approach Nmin-sparse generalized shared control.

Pose estimation is central to several robotics applications such as registration, hand-eye calibration, SLAM, etc. Online pose estimation methods typically use Gaussian distributions to describe the uncertainty in the pose parameters. Such a description can be inadequate when using parameters such as unit-quaternions that are not unimodally distributed. A Bingham distribution can effectively model the uncertainty in unit-quaternions, as it has antipodal symmetry and is defined on a unit-hypersphere. A combination of Gaussian and Bingham distributions is used to develop a linear filter that accurately estimates the distribution of the pose parameters, in their true space. To the best of our knowledge our approach is the first implementation to use a Bingham distribution for 6 DoF pose estimation. Experiments assert that this approach is robust to initial estimation errors as well as sensor noise. Compared to state of the art methods, our approach takes fewer iterations to converge onto the correct pose estimate. The efficacy of the formulation is illustrated with a number of simulated examples on standard datasets as well as real-world experiments.

We introduce Bayesian Eigenobjects (BEOs), a novel object representation that is the first technique able to perform joint classification, pose estimation, and 3D geometric completion on previously unencountered and partially observed query objects. BEOs employ Variational Bayesian Principal Component Analysis (VBPCA) directly on 3D object representations to create generative and compact probabilistic models for classes of 3D objects. Using only depth information, we significantly outperform the current state-of-the-art method for joint classification and 3D completion in both accuracy and query time. Additionally, we show that BEOs are well suited for the extremely challenging task of joint classification, completion, and pose estimation on a large dataset of household objects.

A fundamental requirement for legged robots is to maintain balance and prevent potentially damaging falls whenever possible. As a response to outside disturbances, fall prevention can be achieved by a combination of active balancing actions, e.g. through ankle torques and upper-body motion, and through reactive step placement. While it is widely accepted that stepping is required to respond to large disturbances, the limits of active motions on balancing and step recovery are only well understood for the simplest of walking models. Recent advances in convex optimization-based verification and control techniques enable a more complete understanding of the limits and capabilities of more complex models. In this work, we present an algorithmic approach for formal analysis of the viable-capture basins of walking robots, calculating both inner and outer approximations and corresponding push recovery control strategies. Extending beyond the classic Linear Inverted Pendulum Model (LIPM), we analyze a series of centroidal momentum based planar walking models, examining the effects of center of mass height, angular momentum, and impact dynamics during stepping on capturability. This formal analysis enables an explicit calculation of the differences between these models, and assessment of whether the simplest models ultimately sacrifice capability, and thus stability, when designing push recovery control policies.

In highly constrained settings, e.g., a tentacle-like medical robot maneuvering through narrow cavities in the body for minimally invasive surgery, it may be difficult or impossible for a robot with a generic kinematic design to reach all desirable targets while avoiding obstacles. We introduce a design optimization method to compute kinematic design parameters that enable a single robot to reach as many desirable goal regions as possible while avoiding obstacles in an environment. We focus on the kinematic design of piecewise cylindrical robots, robotic manipulators whose shape can be modeled via cylindrical components. Our method appropriately integrates sampling-based motion planning in configuration space into stochastic optimization in design space so that, over time, our evaluation of a design's ability to reach goals increases in accuracy and our selected designs approach global optimality. We prove the asymptotic optimality of our method and demonstrate performance in simulation for (i) a serial manipulator and (ii) a concentric tube robot, a tentacle-like medical robot that can bend around anatomical obstacles to safely reach clinically-relevant goal regions.

For co-manipulation involving humans and robots, robot controllers that are based on human-human behavior should allow more comfortable and coordinated movement between the human-robot dyad. In this paper, we describe an experiment between human-human dyads where we recorded the force and motion data as leader-follower dyads moved in translation and rotation. The force/motion data was then analyzed for patterns found during lateral translation only. For extended objects, lateral translation and in-place rotation are ambiguous, but this paper determines a way to characterize lateral translation triggers for future use in human-robot interaction. The study has 4 main results. First, interaction forces are non-negligible and are necessary for co-manipulation. Second, minimum-jerk trajectories are found in the lateral direction only for lateral movement. Third, the beginning of a lateral movement is characterized by distinct force triggers by the leader. Fourth, there are different metrics that can be calculated to determine which dyads moved most effectively in the lateral direction.

In this work we present a novel, inductance-based system to measure and control the motion of bellows-driven continuum joints in soft robots. The sensing system relies on coils of wire wrapped around the minor diameters of each bellows on the joint. As the bellows extend, these coils of wire become more distant, decreasing their mutual inductance. Measuring this change in mutual inductance allows us to measure the motion of the joint. By dividing the sensing of the joint into two sections and measuring the motion of each section independently, we are able to measure the overall deformation of the joint with a piece-wise constant-curvature approximation. This technique allows us to measure lateral displacements that would be otherwise unobservable. When measuring bending, the inductance sensors measured the joint orientation with an RMS error of 1.1 degrees. The inductance sensors were also successfully used as feedback to control the orientation of the joint. The sensors proposed and tested in this work provided accurate motion feedback that would be difficult to achieve robustly with other sensors. This sensing system enables the creation of robust, self-sensing, and soft robots based on bellows-driven continuum joints.

In the adaptive information gathering problem, a policy is required to select an informative sensing location using the history of measurements acquired thus far. While there is an extensive amount of prior work investigating effective practical approximations using variants of Shannon's entropy, the efficacy of such policies heavily depends on the geometric distribution of objects in the world. On the other hand, the principled approach of employing online POMDP solvers is rendered impractical by the need to explicitly sample online from a posterior distribution of world maps. We present a novel data-driven imitation learning framework to efficiently train information gathering policies. The policy imitates a clairvoyant oracle - an oracle that at train time has full knowledge about the world map and can compute maximally informative sensing locations. We analyze the learnt policy by showing that offline imitation of a clairvoyant oracle is implicitly equivalent to online oracle execution in conjunction with posterior sampling. This observation allows us to obtain powerful near-optimality guarantees for information gathering problems possessing an adaptive sub-modularity property. As demonstrated on a spectrum of 2D and 3D exploration problems, the trained policies enjoy the best of both worlds - they adapt to different world map distributions while being computationally inexpensive to evaluate.

Our goal is to efficiently learn reward functions encoding a human's preferences for how a dynamical system should act. There are two challenges with this. First, in many problems it is difficult for people to provide demonstrations of the desired system trajectory (like a high-DOF robot arm motion or an aggressive driving maneuver), or to even assign how much numerical reward an action or trajectory should get. We build on work in label ranking and propose to learn from preferences (or comparisons) instead: the person provides the system a relative preference between two trajectories. Second, the learned reward function strongly depends on what environments and trajectories were experienced during the training phase. We thus take an active learning approach, in which the system decides on what preference queries to make. A novel aspect of our work is the complexity and continuous nature of the queries: continuous trajectories of a dynamical system in environments with other moving agents (humans or robots). We contribute a method for actively synthesizing queries that satisfy the dynamics of the system. Further, we learn the reward function from a continuous hypothesis space by maximizing the volume removed from the hypothesis space by each query. We assign weights to the hypothesis space in the form of a log-concave distribution and provide a bound on the number of iterations required to converge. We show that our algorithm converges faster to the desired reward compared to approaches that are not active or that do not synthesize queries in an autonomous driving domain. We then run a user study to put our method to the test with real people.

Humans can ground natural language commands to tasks at both abstract and fine-grained levels of specificity. For instance, a human forklift operator can be instructed to perform a high-level action, like 'grab a pallet' or a low-level action like 'tilt back a little bit.' While robots are also capable of grounding language commands to tasks, previous methods implicitly assume that all commands and tasks reside at a single, fixed level of abstraction. Additionally, methods that do not use multiple levels of abstraction encounter inefficient planning and execution times as they solve tasks at a single level of abstraction with large, intractable state-action spaces closely resembling real world complexity. In this work, by grounding commands to all the tasks or subtasks available in a hierarchical planning framework, we arrive at a model capable of interpreting language at multiple levels of specificity ranging from coarse to more granular. We show that the accuracy of the grounding procedure is improved when simultaneously inferring the degree of abstraction in language used to communicate the task. Leveraging hierarchy also improves efficiency: our proposed approach enables a robot to respond to a command within one second on 90% of our tasks, while baselines take over twenty seconds on half the tasks. Finally, we demonstrate that a real, physical robot can ground commands at multiple levels of abstraction allowing it to efficiently plan different subtasks within the same planning hierarchy.

Neuromorphic electronic systems exhibit advantageous characteristics, in terms of low energy consumption and low response latency, which can be useful in robotic applications that require compact and low power embedded computing resources. However, these neuromorphic circuits still face significant limitations that make their usage challenging: these include low precision, variability of components, sensitivity to noise and temperature drifts, as well as the currently limited number of neurons and synapses that are typically emulated on a single chip. In this paper, we show how it is possible to achieve functional robot control strategies using a mixed signal analog/digital neuromorphic processor interfaced to a mobile robotic platform equipped with an event-based dynamic vision sensor. We provide a proof of concept implementation of obstacle avoidance and target acquisition using biologically plausible spiking neural networks directly emulated by the neuromorphic hardware. To our knowledge, this is the first demonstration of a working spike-based neuromorphic robotic controller in this type of hardware which illustrates the feasibility, as well as limitations, of this approach.

Based on the convex force-motion polynomial model for quasi-static sliding, we derive the kinematic contact model to determine the contact modes and instantaneous object motion on a supporting surface given a position controlled manipulator. The inherently stochastic object-to-surface friction distribution is modelled by sampling physically consistent parameters from appropriate distributions, with only one parameter to control the amount of noise. Thanks to the high fidelity and smoothness of convex polynomial models, the mechanics of patch contact is captured while being computationally efficient without mode selection at support points. The motion equations for both single and multiple frictional contacts are given. Simulation based on the model is validated with robotic pushing and grasping experiments.