RSS 2017 Papers

We present a methodology for fast prototyping of morphologies and controllers for robot locomotion. Going beyond simulation-based approaches, we argue that the form and function of a robot, as well as their interplay with real-world environmental conditions are critical. Hence, fast design and learning cycles are necessary to adapt robot shape and behavior to their environment. To this end, we present a combination of laminate robot manufacturing and sample-efficient reinforcement learning. We leverage this methodology to conduct an extensive robot learning experiment. Inspired by locomotion in sea turtles, we design a low-cost crawling robot with variable, interchangeable fins. Learning is performed using both bio-inspired and original fin designs in an artificial indoor environment as well as a natural environment in the Arizona desert. The findings of this study show that static policies developed in the laboratory do not translate to effective locomotion strategies in natural environments. In contrast to that, sample-efficient reinforcement learning can help to rapidly accommodate changes in the environment or the robot.

Unlike traditional third-person cameras mounted on robots, a first-person camera, captures a person's visual sensorimotor object interactions from up close. In this paper, we study the tight interplay between our momentary visual attention and motor action with objects from a first-person camera. We propose a concept of action-objects---the objects that capture person's conscious visual (watching a TV) or tactile (taking a cup) interactions. Action-objects may be task-dependent but since many tasks share common person-object spatial configurations, action-objects exhibit a characteristic 3D spatial distance and orientation with respect to the person. We design a predictive model that detects action-objects using EgoNet, a joint two-stream network that holistically integrates visual appearance (RGB) and 3D spatial layout (depth and height) cues to predict per-pixel likelihood of action-objects. Our network also incorporates a first-person coordinate embedding, which is designed to learn a spatial distribution of the action-objects in the first-person data. We demonstrate EgoNet's predictive power, by showing that it consistently outperforms previous baseline approaches. Furthermore, EgoNet also exhibits a strong generalization ability, i.e., it predicts semantically meaningful objects in novel first-person datasets. Our method's ability to effectively detect action-objects could be used to improve robots' understanding of human-object interactions.

This paper derives nonlinear feedback control synthesis for general control affine systems using second-order actions---the needle variations of optimal control---as the basis for choosing each control response to the current state. A second result of the paper is that the method provably exploits the nonlinear controllability of a system by virtue of an explicit dependence of the second-order needle variation on the Lie bracket between vector fields. As a result, each control decision necessarily decreases the objective when the system is nonlinearly controllable using first-order Lie brackets. Simulation results using a differential drive cart, an underactuated kinematic vehicle in three dimensions, and an underactuated dynamic model of an underwater vehicle demonstrate that the method finds control solutions when the first-order analysis is singular. Moreover, the simulated examples demonstrate superior convergence when compared to synthesis based on first-order needle variations. Lastly, the underactuated dynamic underwater vehicle model demonstrates the convergence even in the presence of a velocity field.

Executing agile quadrotor maneuvers with cable-suspended payloads is a challenging problem and complications induced by the dynamics typically require trajectory optimization. State-of-the-art approaches often need significant computation time and complex parameter tuning. We present a novel dynamical model and a fast trajectory optimization algorithm for quadrotors with a cable-suspended payload. Our first contribution is a new formulation of the suspended payload behavior, modeled as a link attached to the quadrotor with a combination of two revolute joints and a prismatic joint, all being passive. Differently from state of the art, we do not require the use of hybrid modes depending on the cable tension. Our second contribution is a fast trajectory optimization technique for the aforementioned system. Our model enables us to pose the trajectory optimization problem as a Mathematical Program with Complementarity Constraints (MPCC). Desired behaviors of the system (e.g., obstacle avoidance) can easily be formulated within this framework. We show that our approach outperforms the state of the art in terms of computation speed and guarantees feasibility of the trajectory with respect to both the system dynamics and control input saturation, while utilizing far fewer tuning parameters. We experimentally validate our approach on a real quadrotor showing that our method generalizes to a variety of tasks, such as flying through desired waypoints while avoiding obstacles, or throwing the payload toward a desired target. To the best of our knowledge, this is the first time that three-dimensional, agile maneuvers exploiting the system dynamics have been achieved on quadrotors with a cable-suspended payload.

We present an extension to Experience-driven Predictive Control (EPC) that leverages a Gaussian belief propagation strategy to compute an uncertainty set bounding the evolution of the system state in the presence of time-varying state uncertainty. This uncertainty set is used to tighten the constraints in the predictive control formulation via a chance constrained approach, thereby providing a probabilistic guarantee of constraint satisfaction. The parameterized form of the controllers produced by EPC coupled with online uncertainty estimates ensures this robust constraint satisfaction property persists even as the system switches controllers and experiences variations in the uncertainty model. We validate the online performance and robust constraint satisfaction of the proposed Robust EPC algorithm through a series of experimental trials with a small quadrotor platform subjected to changes in state estimate quality.

This paper explores the quasi-static motion of a planar slider being pushed or pulled through a single contact point assumed not to slip. The main contribution is to derive a method for computing exact bounds on the object's motion for classes of pressure distributions where the center of pressure is known but the distribution of support forces is unknown. The second contribution is to show that the exact motion bounds can be used to plan robotic pulling trajectories that guarantee convergence to the final pose. The planner was tested on the task of pulling an acrylic rectangle to random locations within the robot workspace. The generated plans were accurate to 4.00mm +- 3.02mm of the target position and 4.35 degrees +- 3.14 degrees of the target orientation.

This paper presents a tool for addressing a key component in many algorithms for planning robot trajectories under uncertainty: evaluation of the safety of a robot whose actions are governed by a closed-loop feedback policy near a nominal planned trajectory. We describe an adaptive importance sampling Monte Carlo framework that enables the evaluation of a given control policy for satisfaction of a probabilistic collision avoidance constraint which also provides an associated certificate of accuracy (in the form of a confidence interval). In particular this adaptive technique is well-suited to addressing the complexities of rigid-body collision checking applied to non-linear robot dynamics. As a Monte Carlo method it is amenable to parallelization for computational tractability, and is generally applicable to a wide gamut of simulatable systems, including alternative noise models. Numerical experiments demonstrating the effectiveness of the adaptive importance sampling procedure are presented and discussed.

The overarching goal of this work is to efficiently enable end-users to correctly anticipate a robot's behavior in novel situations. Since a robot's behavior is often a direct result of its underlying objective function, our insight is that end-users need to have an accurate mental model of this objective function in order to understand and predict what the robot will do. While people naturally develop such a mental model over time through observing the robot act, this familiarization process may be lengthy. Our approach reduces this time by having the robot model how people infer objectives from observed behavior, and then it selects those behaviors that are maximally informative. The problem of computing a posterior over objectives from observed behavior is known as Inverse Reinforcement Learning (IRL), and has been applied to robots learning human objectives. We consider the problem where the roles of human and robot are swapped. Our main contribution is to recognize that unlike robots, humans will not be exact in their IRL inference. We thus introduce two factors to define candidate approximate-inference models for human learning in this setting, and analyze them in a user study in the autonomous driving domain. We show that certain approximate-inference models lead to the robot generating example behaviors that better enable users to anticipate what it will do in novel situations. Our results also suggest, however, that additional research is needed in modeling how humans extrapolate from examples of robot behavior.

This work addresses the problem of efficient online exploration and mapping using multi-robot teams via a distributed algorithm for planning for multi-robot exploration---distributed sequential greedy assignment (DSGA)---based on the sequential greedy assignment (SGA) algorithm. SGA permits bounds on suboptimality but requires all robots to plan in series. Rather than plan for robots sequentially as in SGA, DSGA assigns plans to subsets of robots during a fixed number of rounds. DSGA retains the same suboptimality bounds as SGA with the addition of a term describing suboptimality introduced due to redundant sensor information. We use this result to extend a single-robot planner based on Monte-Carlo tree search to the multi-robot domain and evaluate the resulting planner in simulated exploration of a confined and cluttered environment. The experimental results show that suboptimality due to redundant sensor information introduced by the distributed planning rounds remains near zero in practice when using as few as two or three distributed planning rounds and that DSGA achieves similar or better objective values and entropy reduction as SGA while providing a 2--6 times computational speedup for multi-robot teams ranging from 4 to 32 robots.

A novel method is presented for efficiently testing the stability of an object under gravity and contact forces, that accommodates empirical determination of the set of admissible forces exerted at contacting pairs of surfaces. These admissible force volumes may exhibit a wide variety of geometries, including anisotropy, adhesion, and even non-convexity. The method discretizes the contact region into patches, performs a convex decomposition of a polyhedral approximation to each admissible force volume, and then formulates the problem as a mixed integer linear program. The model can also accommodate articulated robot hands with joint torques, joint frictions, and spring preloads. Predictions of our method are evaluated experimentally in object lifting tasks using a gripper that exploits microspines to exert strongly anisotropic forces.

We study the effectiveness of metrics for Multi-Robot Motion-Planning (MRMP) when using RRT-style sampling-based planners. These metrics play the crucial role of determining the nearest neighbors of configurations and in that they regulate the connectivity of the underlying roadmaps produced by the planners and other properties like the quality of solution paths. After screening over a dozen different metrics we focus on the five most promising ones' two more traditional metrics, and three novel ones which we propose here, adapted from the domain of shape-matching. In addition to the novel multi-robot metrics, a central contribution of this work are tools to analyze and predict the effectiveness of metrics in the MRMP context. We identify a suite of possible substructures in the configuration space, for which it is fairly easy (i) to define a so-called natural distance, which allows us to predict the performance of a metric. This is done by comparing the distribution of its values for sampled pairs of configurations to the distribution induced by the natural distance; (ii) to define equivalence classes of configurations and test how well a metric covers the different classes. We provide experiments that attest to the ability of our tools to predict the effectiveness of metrics: those metrics that qualify in the analysis yield higher success rate of the planner with fewer vertices in the roadmap. We also show how combining several metrics together leads to better results (success rate and size of roadmap) than using a single metric.

An inspiration for developing a bipedal walking system is the ability to navigate rough terrain with discrete footholds like stepping stones. In this paper, we present a novel methodology to overcome the problem of dynamic walking over stepping stones with significant random changes to step length and step height at each step. Using a 2-step gait optimization, we not only consider the desired location of the next footstep but also the current configuration of the robot, thereby resolving the problem of step transition when we switch between different walking gaits. We then use gait interpolation to generate the desired walking gait in real-time. We demonstrate the method on a planar dynamical walking model of ATRIAS, an underactuated bipedal robot walking over a randomly generated stepping stones with step length and step height changing in the range of [30:80] (cm) and [-30:30] (cm) respectively. Experimental validation on the real robot was also successful for the problem of dynamic walking on stepping stones with step lengths varied within [23:78] (cm) and average walking speed of 0.6 (m/s).

In this paper, we extend the concept of control barrier functions, developed initially for continuous time systems, to the discrete-time domain. We demonstrate safety-critical control for nonlinear discrete-time systems with applications to 3D bipedal robot navigation. Particularly, we mathematically analyze two different formulations of control barrier functions, based on their continuous-time counterparts, and demonstrate how these can be applied to discrete-time systems. We show that the resulting formulation is a nonlinear program in contrast to the quadratic program for continuous-time systems and under certain conditions, the nonlinear program can be formulated as a quadratically constrained quadratic program. Furthermore, using the developed concept of discrete control barrier functions, we present a novel control method to address the problem of navigation of a high-dimensional bipedal robot through environments with moving obstacles that present time-varying safety-critical constraints.

Developing robots capable of making sense of their environment requires the ability to learn from observations. An important paradigm that allows for robots to both imitate humans and gain an understanding of the tasks people perform is that of action primitive discovery. Action primitives have been used as a representation of the main building blocks that compose motion. Automatic primitive discovery is an active area of research, with existing methods that can provide viable solutions for learning primitives from demonstrations. However, when we learn primitives directly from raw data, we need a mechanism to determine those primitives that are appropriate for the task at hand: is "brushing one's teeth" a suitable primitive or are the actions of "grabbing the toothbrush", "adding toothpaste onto it", and "executing the brushing motion" better suited? It is this level of granularity that is important for determining well-suited primitives for applications. Existing methods for learning primitives do not provide a solution for discovering their granularity. Rather, these techniques stop at arbitrarily chosen levels, and often use clear, repetitive actions in order to easily label the primitives. Our contribution provides a framework for discovering the appropriate granularity level of learned primitives for a task. We apply our framework to action primitives learned from a set of motion capture data obtained from human demonstrations that includes hand and object motions. This helps find a well-suited granularity level for our task, avoiding the use of low levels that don't capture the necessary core pattern in the actions, or high levels that miss important differences between actions. Our results show that this framework is able to discover the best suited primitive granularity level for a specific application.

To reduce data collection time for deep learning of robust robotic grasp plans, we explore training from a synthetic dataset of 6.7 million point clouds, grasps, and robust analytic grasp metrics generated from thousands of 3D models from Dex-Net 1.0 in randomized poses on a table. We use the resulting dataset, Dex-Net 2.0, to train a Grasp Quality Convolutional Neural Network (GQ-CNN) model that rapidly predicts the probability of success of grasps from depth images, where grasps are specified as the planar position, angle, and depth of a gripper relative to an RGB-D sensor. Experiments with over 1,000 trials on an ABB YuMi comparing grasp planning methods on singulated objects suggest that a GQ-CNN trained with only synthetic data from Dex-Net 2.0 can be used to plan grasps in 0.8s with a success rate of 93% on eight known objects with adversarial geometry and is 3x faster than registering point clouds to a precomputed dataset of objects and indexing grasps. The Dex-Net 2.0 grasp planner is also the highest performing method on a dataset of 10 novel rigid objects and achieves 99% precision (one false positive out of 69 grasps classified as robust) on a dataset of 40 novel household objects, some of which are articulated or deformable.

This paper presents the design and optimization of a self-adaptive, a.k.a. underactuated, finger targeted to be used with collaborative robots. Typical robots, whether collaborative or not, mostly rely on standard translational grippers for pick-and-place operations. These grippers are constituted from an actuated motion platform on which a set of jaws is rigidly attached. These jaws are often designed to secure a precise and limited range of objects through the application of pinching forces. In this paper, the design of a self-adaptive robotic finger is presented which can be attached to these typical translational gripper to replace the common monolithic jaws and provide the gripper with shape-adaptation capabilities without any control or sensors. A new design is introduced here and specially optimized for collaborative robots. The kinetostatic analysis of this new design is briefly discussed and then followed by the optimization of relevant geometric parameters. Finally, a practical prototype attached to a very common collaborative robot is demonstrated. While the resulting finger design could be attached to any translational gripper, specifically targeting collaborative robots as an application comes with certain specificities in the choice of the design parameters as will be shown and the optimized parameters that are found match them.

In this paper, using the Hypercube Diagonal Experiment we first investigate the convergence rates of sampling-based path-planning algorithms in terms of the dimensionnality of the search space. We show that the probability of sampling a point that improves the solution decreases exponentially with the dimension of the problem. We then analyze how the samples can be repositioned in the search space in order to minimize the approximation error. Finally, we present the DRRT (Deformable Rapidly Exploring Random Tree) algorithm that utilizes optimization of sample location in the framework of RRT algorithms to improve convergence. It is shown that the DRRT algorithm significantly outperforms all current sampling-based algorithms in terms of convergence.

Many critical robotics applications require robustness to disturbances arising from unplanned forces, state uncertainty, and model errors. Motion planning algorithms that explicitly reason about robustness require a coupling of trajectory optimization and feedback design, where the system's closed-loop response to bounded disturbances is optimized. Due to the often-heavy computational demands of solving such problems, the practical application of robust trajectory optimization in robotics has so far been limited. We derive a tractable robust optimization algorithm that combines direct transcription with linear-quadratic control design to reason about closed-loop responses to disturbances. In the case of ellipsoidal disturbance sets, the state and input deviations along a nominal trajectory can be computed locally in closed form, thus allowing for fast evaluations of robust cost and constraint functions. The resulting algorithm, called DIRTREL, is an extension of classical direct transcription that demonstrably improves tracking performance over non-robust formulations while incurring only a modest increase in computational cost. We evaluate the algorithm in several simulated robot control tasks.

3D scene understanding is important for robots to interact with the 3D world in a meaningful way. Most previous works on 3D scene understanding focus on recognizing geometrical or semantic properties of the scene independently. In this work, we introduce Data Associated Recurrent Neural Networks (DA-RNNs), a novel framework for joint 3D scene mapping and semantic labeling. DA-RNNs use a new recurrent neural network architecture for semantic labeling on RGB-D videos. The output of the network is integrated with mapping techniques such as KinectFusion in order to inject semantic information into the reconstructed 3D scene. Experiments conducted on a real world dataset and a synthetic dataset with RGB-D videos demonstrate the ability of our method in semantic 3D scene mapping.