RSS 2017 论文列表

Autonomous navigation of miniaturized robots (e.g., nano/pico aerial vehicles) is currently a grand challenge for robotics research, due to the need for processing a large amount of sensor data (e.g., camera frames) with limited on-board computational resources. In this paper we focus on the design of a visual-inertial odometry (VIO) system in which the robot estimates its ego-motion (and a landmark-based map) from on- board camera and IMU data. We argue that scaling down VIO to miniaturized platforms (without sacrificing performance) requires a paradigm shift in the design of perception algorithms, and we advocate a co-design approach in which algorithmic and hardware design choices are tightly coupled. Our contribution is four-fold. First, we discuss the VIO co-design problem, in which one tries to attain a desired resource-performance trade-off, by making suitable design choices (in terms of hardware, algorithms, implementation, and parameters). Second, we characterize the design space, by discussing how a relevant set of design choices affects the resource-performance trade-off in VIO. Third, we provide a systematic experiment-driven way to explore the design space, towards a design that meets the desired trade-off. Fourth, we demonstrate the result of the co-design process by providing a VIO implementation on specialized hardware and showing that such implementation has the same accuracy and speed of a desktop implementation, while requiring a fraction of the power.

Reward function design and exploration time are arguably the biggest obstacles to the deployment of reinforcement learning (RL) agents in the real world. In many real-world tasks, designing a reward function takes considerable hand engineering and often requires additional and potentially visible sensors to be installed just to measure whether the task has been executed successfully. Furthermore, many interesting tasks consist of multiple implicit intermediate steps that must be executed in sequence. Even when the final outcome can be measured, it does not necessarily provide feedback on these intermediate steps or sub-goals. To address these issues, we propose leveraging the abstraction power of intermediate visual representations learned by deep models to quickly infer perceptual reward functions from small numbers of demonstrations. We present a method that is able to identify key intermediate steps of a task from only a handful of demonstration sequences, and automatically identify the most discriminative features for identifying these steps. This method makes use of the features in a pre-trained deep model, but does not require any explicit specification of sub-goals. The resulting reward functions, which are dense and smooth, can then be used by an RL agent to learn to perform the task in real-world settings. To evaluate the learned reward functions, we present qualitative results on two real-world tasks and a quantitative evaluation against a human-designed reward function. We also demonstrate that our method can be used to learn a complex real-world door opening skill using a real robot, even when the demonstration used for reward learning is provided by a human using their own hand. To our knowledge, these are the first results showing that complex robotic manipulation skills can be learned directly and without supervised labels from a video of a human performing the task. Supplementary material and dataset are available at sermanet.github.io/rewards

This paper discusses several design issues pertaining to the design of anthropomorphic underactuated robotic or prosthetic hands. The design of the underactuation mechanism that drives the fingers is first addressed. A double-stage mechanism is proposed that includes a lever arm to couple the motion of the thumb and the fingers as well as an additional pulley that amplifies the input force. Then, the design of the underactuated fingers is discussed and a novel rotational sliding joint is proposed. The geometric and mechanical design of a thumb mechanism is also presented that allows the reconfiguration of the thumb in the three most favorable poses. Finally, a prototype is presented that implements the design choices discussed in the paper and its effectiveness is qualitatively discussed based on a series of experiments.

We address the problem of uncertainty-aware local collision avoidance within the context of time-to-collision based navigation of multiple agents. We consider two specific models that account for uncertainty in the future trajectories of interacting agents: an isotropic model which conservatively considers all possible errors, and an adversarial model that assumes the error is towards a head-on collision. We compare the two models experimentally via a number of simulation scenarios, and also provide theoretical guarantees about the collision avoidance behavior of the agents.

Trajectory generation approaches for mobile robots generally aim to optimize with respect to a cost function such as energy, execution time, or other mission-relevant parameters within the constraints of vehicle dynamics and obstacles in the environment. We propose to add the cost of state observability to the trajectory optimization in order to ensure fast and accurate state estimation throughout the mission while still respecting the constraints of vehicle dynamics and the environment. Our approach finds a dynamically feasible estimation-optimized trajectory in a sequence of connected convex polytopes representing free space in the environment. In addition, we show a statistical procedure that enables observability-aware trajectory optimization for heterogeneous states in the system both in magnitude and units, which was not supported in previous formulations. We validate our approach with extensive simulations of a visual-inertial state estimator on an aerial platform as a specific realization of our general method. We show that the optimized trajectories lead to more accurate navigation while eliminating the need for a separate calibration procedure.

Robust robotic perception and manipulation of household objects requires the ability to detect, localize and manipulate a wide variety of objects, which may be mirror reflective like polished metal, glossy like smooth plastic, or transparent like glass; for example, picking a metal fork out of a sink full of running water or screwing a metal nut onto a bolt. Existing perceptual approaches based on photographs only take into account the average intensity of light arriving at each pixel from one direction, which limits their ability to account for these non-Lambertian scenes. To address this problem, we demonstrate time-lapse light field photography with an eye-in-hand camera of a manipulator robot. An eye-in-hand robot can capture both the intensity of rays, as in a conventional photograph, as well as the direction of the rays. We present a formal model for robotic light-field photography that fits into a probabilistic robotics framework. Using this model, we can synthesize orthographic photographs, remove specular highlights from those photographs, and perform 3D reconstruction with a monocular camera by finding approximate maximum-likelihood estimates. This information can be used to detect, localize and manipulate non-Lambertian objects in non-Lambertian scenes: our approach enables the Baxter robot to pick a shiny metal fork out of a sink filled with running water 24/25 times, as well as to localize objects well enough to screw a nut onto a quarter inch bolt. The techniques in this paper point the way toward new approaches to robotic perception that leverage a robot's ability to move its camera to infer the state of the external world.

Being inherently compliant, the robotic artificial muscles are increasingly popular in applications such as safe human-robot interaction, legged robotics, prostheses and orthoses, and soft robotics. Their full utilization is often challenged by the coupled hysteresis among input, strain, and tension force. Although conventional two-dimensional hysteresis models are available, no prior studies on three-dimensional hysteresis models with coupled inputs have been reported for robotic artificial muscles. This paper presents a new approach to capturing the three-dimensional hysteresis of robotic artificial muscles by embedding a two-stage Preisach model. The proposed method is applied to shape memory alloy (SMA) actuators. Since direct temperature measurement of the SMA actuator is not available, the concept of temperature surrogate, representing the constant voltage value in Joule heating that would result in a given temperature at the steady-state, is adopted. The proposed approach is utilized to capture the hysteresis among temperature surrogate, contraction length, and force of an SMA actuator. Model verification is further conducted. For comparison purposes, two modeling approaches, namely, the Summed Preisach and the Linear Preisach, are also realized. Experimental results demonstrate that the proposed scheme can effectively characterize and estimate the three-dimensional hysteresis in SMA actuators. This study can be applied towards other robotic artificial muscles such as McKibben actuators and Super-coiled Polymer actuators.

Detection of objects in cluttered indoor environments is one of the key enabling functionalities for service robots. The best performing object detection approaches in computer vision exploit deep Convolutional Neural Networks (CNN) to simultaneously detect and categorize the objects of interest in cluttered scenes. Training of such models typically requires large amounts of annotated training data which is time consuming and costly to obtain. In this work we explore the ability of using synthetically generated composite images for training state-of-the-art object detectors, especially for object instance detection. We superimpose 2D images of textured object models into images of real environments at variety of locations and scales. Our experiments evaluate different superimposition strategies ranging from purely image-based blending all the way to depth and semantics informed positioning of the object models into real scenes. We demonstrate the effectiveness of these object detector training strategies on two publicly available datasets, the GMU-Kitchens and the Washington RGB-D Scenes v2. As one observation, augmenting some hand-labeled training data with synthetic examples carefully composed onto scenes yields object detectors with comparable performance to using much more hand-labeled data. Broadly, this work charts new opportunities for training detectors for new objects by exploiting existing object model repositories in either a purely automatic fashion or with only a very small number of human-annotated examples.

Design and control of a novel extra robotic arm attached to the shoulder of a worker for performing tasks in the overhead area are presented. The wearable robot, called Supernumerary Robotic Limbs (SRL), can lift an object and hold it while the wearer is securing the object using a tool with both hands. The worker does not have to take a laborious posture for a long time, reducing fatigue and injuries. Furthermore, a single worker can execute the task, which would otherwise require two workers. Two technical challenges and novel solutions are presented. One is to make the wearable robot simple and lightweight with use of a new type of granular jamming gripper that can grasp diverse objects from an arbitrary direction. This eliminates the need for orienting the gripper against the object with three-axis wrist joints, reducing the number of degrees of freedom (DOF) from 6 to 3. The other is an effective control algorithm that allows the wearer to move freely while the robot on the shoulder is holding an object. Unlike a robot sitting on a floor, the SRL worn by a human is disturbed by the movement of the wearer. An admittance-based control algorithm allows the robot to hold the object stably and securely despite the human movement and changes in posture. A 3 DOF prototype robot with a new granular jamming gripper and an ergonomic body mounting gear is developed and tested. It is demonstrated that the robot can hold a large object securely in the overhead area despite the movement of the wearer while performing an assembly work.

We provide a unified probabilistic framework for trajectory estimation and planning. The key idea is to view these two problems, usually considered separately, as a single problem. At each time-step the robot is tasked with finding the complete continuous-time trajectory from start to goal. This can be quite difficult; the robot must contend with a potentially high-degree-of-freedom (DOF) trajectory space, uncertainty due to limited sensing capabilities, model inaccuracy, and the stochastic effect of executing actions, and the robot must find the solution in (faster than) real time. To overcome these challenges, we build on recent probabilistic inference approaches to continuous-time localization and mapping and continuous-time motion planning. We solve the joint problem by iteratively recomputing the maximum a posteriori trajectory conditioned on all available sensor data and cost information. Finally, we evaluate our framework empirically in both simulation and on a mobile manipulator.

There has been a great deal of progress in developing probabilistically complete methods that move beyond motion planning to multi-modal problems including various forms of task planning. This paper presents a general-purpose formulation of a large class of discrete-time planning problems, with hybrid state and action spaces. The formulation characterizes conditions on the submanifolds in which solutions lie, leading to a characterization of robust feasibility that incorporates dimensionality-reducing constraints. It then connects those conditions to corresponding conditional samplers that are provided as part of a domain specification. We present domain-independent sample-based planning algorithms and show that they are both probabilistically complete and computationally efficient on a set of challenging benchmark problems.

Robots that use learned perceptual models in the real world must be able to safely handle cases where they are forced to make decisions in scenarios that are unlike any of their training examples. However, state-of-the-art deep learning methods are known to produce erratic or unsafe predictions when faced with novel inputs. Furthermore, recent ensemble, bootstrap and dropout methods for quantifying neural network uncertainty may not efficiently provide accurate uncertainty estimates when queried with inputs that are very different from their training data. Rather than unconditionally trusting the predictions of a neural network for unpredictable real-world data, we use an autoencoder to recognize when a query is novel, and revert to a safe prior behavior. With this capability, we can deploy an autonomous deep learning system in arbitrary environments, without concern for whether it has received the appropriate training. We demonstrate our method with a vision-guided robot that can leverage its deep neural network to navigate 50% faster than a safe baseline policy in familiar types of environments, while reverting to the prior behavior in novel environments so that it can safely collect additional training data and continually improve. A video illustrating our approach is available at: http://groups.csail.mit.edu/rrg/videos/safe visual navigation.

The literature on Inverse Reinforcement Learning (IRL) typically assumes that humans take actions in order to minimize the expected value of a cost function, i.e., that humans are risk neutral. Yet, in practice, humans are often far from being risk neutral. To fill this gap, the objective of this paper is to devise a framework for risk-sensitive IRL in order to explicitly account for an expert's risk sensitivity. To this end, we propose a flexible class of models based on coherent risk metrics, which allow us to capture an entire spectrum of risk preferences from risk-neutral to worst-case. We propose efficient algorithms based on Linear Programming for inferring an expert's underlying risk metric and cost function for a rich class of static and dynamic decision-making settings. The resulting approach is demonstrated on a simulated driving game with ten human participants. Our method is able to infer and mimic a wide range of qualitatively different driving styles from highly risk-averse to risk-neutral in a data-efficient manner. Moreover, comparisons of the Risk-Sensitive (RS) IRL approach with a risk-neutral model show that the RS-IRL framework more accurately captures observed participant behavior both qualitatively and quantitatively.

Time series classification is an important task in robotics that is often solved using supervised machine learning. However, classifier models are typically not `readable' in the sense that humans cannot intuitively learn useful information about the relationship between inputs and outputs. In this paper, we address the problem of rich time series classification where we propose a novel framework for finding a temporal logic classifier specified in a human-readable form. The classifier is represented as a signal temporal logic (STL) formula that is expressive in capturing spatial, temporal and logical relations from a continuous-valued dataset over time. In the framework, we first find a set of representative logical formulas from the raw dataset, and then construct an STL classifier using a tree-based clustering algorithm. We show that the framework runs in polynomial time and validate it using simulated examples where our framework is significantly more efficient than the closest existing framework (up to 920 times faster).

Robots designed to interact with humans in realistic environments must be able to handle uncertainty with respect to the identities and properties of the people, places, and things found in their environments. When humans refer to these entities using under-specified language, robots must often generate clarification requests to determine which entities were meant. In this paper, we present recommendations for designers of robots needing to generate such requests, and show how a Dempster-Shafer theoretic pragmatic reasoning component capable of generating requests to clarify pragmatic uncertainty can also generate requests to resolve referential uncertainty when integrated with a probabilistic reference resolution component.

This paper proposes a novel approach to performing in-grasp manipulation planning: the problem of moving an object with reference to the palm from an initial pose to a goal pose without breaking or making contacts. Our method to perform in-grasp manipulation uses kinematic trajectory optimization which requires no knowledge of dynamic properties of the object or the robot. We define a cost function that attempts to maintain the initial grasp points, while relaxing the constraint that the contacts between finger and object remain rigid. Hence, we name this new formulation relaxed-rigidity constraints. We implement our approach on an Allegro robot hand and perform experiments on 10 objects from the YCB dataset. However, the implementation would work for any object the robot can grasp. We perform thorough analysis and compare to alternative optimization formulations. Our method reaches the desired object pose with a median position error of 13mm across all of the 500 trials without ever dropping the object.

Simulators are powerful tools for reasoning about a robot's interactions with its environment. However, when simulations diverge from reality, that reasoning becomes less useful. In this paper, we show how to close the loop between liquid simulation and real-time perception. We use observations of liquids to correct errors when tracking the liquid's state in a simulator. Our results show that closed-loop simulation is an effective way to prevent large divergence between the simulated and real liquid states. As a direct consequence of this, our method can enable reasoning about liquids that would otherwise be infeasible due to large divergences, such as reasoning about occluded liquid.

As drones and autonomous cars become more widespread it is becoming increasingly important that robots can operate safely under realistic conditions. The noisy information fed into real systems means that robots must use estimates of the environment to plan navigation. Efficiently guaranteeing that the resulting motion plans are safe under these circumstances has proved difficult. We examine how to guarantee that a trajectory or policy is safe with only imperfect observations of the environment. We examine the implications of various mathematical formalisms of safety and arrive at a mathematical notion of safety of a long-term execution, even when conditioned on observational information. We present efficient algorithms that can prove that trajectories or policies are safe with much tighter bounds than in previous work. Notably, the complexity of the environment does not affect our method's ability to evaluate if a trajectory or policy is safe. We then use these safety checking methods to design a safe variant of the RRT planning algorithm.

We present a theoretical analysis of a recent whole body motion planning method, the Randomized Possibility Graph, which uses a high-level decomposition of the feasibility constraint manifold in order to rapidly find routes that may lead to a solution. These routes are then examined by lower-level planners to determine feasibility. In this paper, we show that this approach is probabilistically complete for bipedal robots performing quasi-static walking in "semi-unstructured" environments. Furthermore, we show that the decomposition into higher and lower level planners allows for a considerably higher rate of convergence in the probability of finding a solution when one exists. We illustrate this convergence with a series of simulated scenarios.