RSS 2018 Papers

Small uninhabited aerial vehicles (sUAVs) commonly rely on active propulsion to stay airborne, which limits flight time and range. To address this, autonomous soaring seeks to utilize free atmospheric energy in the form of updrafts (thermals). However, their irregular nature at low altitudes makes them hard to exploit for existing methods. We model autonomous thermalling as a POMDP and present a receding-horizon controller based on it. We implement it as part of ArduPlane, a popular open-source autopilot, and compare it to an existing alternative in a series of live flight tests involving two sUAVs thermalling simultaneously, with our POMDP-based controller showing a significant advantage.

We present and validate a property-driven autonomous system that modifies its environment to achieve and maintain navigability over a highly irregular 3-dimensional terrain. In our approach we use decision procedures that tie building actions to the terrain model, giving rise to adaptive and robust building behavior. The building algorithm is driven by continuous evaluation and reaction to terrain properties, rather than relying on a structure blueprint. This capability is essential in robotic systems that operate in unstructured outdoor or remote environments, either on their own or as part of a team. We demonstrate the effectiveness of our approach by running a low-cost robot system that can build with compliant bags in a variety of irregular terrains.

Deep reinforcement learning (RL) has proven a powerful technique in many sequential decision making domains. However, robotics poses many challenges for RL, most notably training on a physical system can be expensive and dangerous, which has sparked significant interest in learning control policies using a physics simulator. While several recent works have shown promising results in transferring policies trained in simulation to the real world, they often do not fully utilize the advantage of working with a simulator. In this work, we propose the Asymmetric Actor Critic, which learns a vision-based control policy while taking advantage of access to the underlying state to significantly speed up training. Concretely, our algorithm employs an actor-critic training algorithm in which the critic is trained on full states while the actor (or policy) is trained on images. We show that using these asymmetric inputs improves performance on a range of simulated tasks. Finally, we combine this method with domain randomization and show real robot experiments for several tasks like picking, pushing, and moving a block. We achieve this simulation to real-world transfer without training on any real-world data. Videos of these experiments can be found in www.goo.gl/b57WTs.

Rigid body dynamics is a well-established framework in robotics. It can be used to expose the analytic form of kinematic and dynamic functions of the robot model. So far, two major algorithms, namely the recursive Newton-Euler algorithm (RNEA) and the articulated body algorithm (ABA), have been proposed to compute the inverse dynamics and the forward dynamics in a few microseconds. Evaluating their derivatives is an important challenge for various robotic applications (optimal control, estimation, co-design or reinforcement learning). However it remains time consuming, whether using finite differences or automatic differentiation. In this paper, we propose new algorithms to efficiently compute them thanks to closed-form formulations. Using the chain rule and adequate algebraic differentiation of spatial algebra, we firstly differentiate explicitly RNEA. Then, using properties about the derivative of function composition, we show that the same algorithm can also be used to compute the derivatives of ABA with a marginal additional cost. For this purpose, we introduce a new algorithm to compute the inverse of the joint-space inertia matrix, without explicitly computing the matrix itself. All the algorithms are implemented in our open-source C++ framework called Pinocchio. Benchmarks show computational costs varying between 3 microseconds (for a 7-dof arm) up to 17 microseconds (for a 36-dof humanoid), outperforming the alternative approaches of the state of the art.

We present an end-to-end imitation learning system for agile, off-road autonomous driving using only low-cost on-board sensors. By imitating a model predictive controller equipped with advanced sensors, we train a deep neural network control policy to map raw, high-dimensional observations to continuous steering and throttle commands. Compared with recent approaches to similar tasks, our method requires neither state estimation nor on-the-fly planning to navigate the vehicle. Our approach relies on, and experimentally validates, recent imitation learning theory. Empirically, we show that policies trained with online imitation learning overcome well-known challenges related to covariate shift and generalize better than policies trained with batch imitation learning. Built on these insights, our autonomous driving system demonstrates successful high-speed off-road driving, matching the state-of-the-art performance.

This paper reports an adaptive sensor bias estimator and attitude observer operating directly on SO(3) for true-North gyrocompass systems that utilize six-degree of freedom inertial measurement units (IMUs) with three-axis accelerometers and three-axis gyroscopes (without magnetometers). Most present-day low-cost robotic vehicles employ attitude estimation systems that employ micro-electromechanical systems (MEMS) magnetometers, angular rate gyros, and accelerometers to estimate magnetic heading and attitude with limited heading accuracy. Present day MEMS gyros are not sensitive enough to dynamically detect Earth's rotation, and thus cannot be used to estimate true-North geodetic heading. In contrast, the reported gyrocompass system utilizes fiber optic gyroscope (FOG) IMU gyro and MEMS accelerometer measurements (without magnetometers) to dynamically estimate the instrument's time-varying attitude in real-time while the instrument is subject to a priori unknown rotations. Stability proofs, preliminary simulations, and a fullscale vehicle trial are reported that suggest the viability of the true-North gyrocompass system to provide dynamic real-time true-North heading, pitch, and roll while utilizing a comparatively low-cost FOG IMU.

Robot grasping of objects based on variable stiffness actuation not only improves the safety and robustness of the grasp but also enhances dynamic manipulation. In this paper, we present the design aspects of a variable stiffness gripper and demonstrate how the controllable compliance of the fingers can improve the performance in dynamic manipulation tasks such as hammering/hitting. The proposed gripper consists of two parallel fingers and repulsive magnets are used as the nonlinear springs between gripper actuators and fingers. The position and force-stiffness characteristics of the fingers are adjusted simultaneously, by controlling the air-gaps between magnets. Finally, the application of the gripper in a nail hammering task is studied as an example of dynamic manipulation. For this purpose, an optimal stiffness control problem is solved to maximize the impact force of the hammering task through maximizing the kinetic energy of the grasped object at the hitting instance. Despite the simplicity of the design, experimental results indicate the effectiveness of the gripper for dynamic manipulation.

We describe a surgical system that autonomously searches for tumors and dynamically displays a computer graphic model of them super-imposed on the organ (or in our case, phantom). Once localized, the phantom is tracked in real time and augmented with overlaid stiffness information in 3D. We believe that such a system has the potential to quickly reveal the location and shape of tumors and the visual overlay will reduce the cognitive overload of the surgeon. The contribution of this paper is the integration of disparate technologies to achieve this system. In fact, to the best of our knowledge, our approach is one of the first to incorporate state-of-the-art methods in registration, force sensing and tumor localization into a unified surgical system. First, the preoperative model is registered to the intra-operative scene using a Bingham distribution-based filtering approach. An active level set estimation is then used to find the location and the shape of the tumors. We use a recently developed miniature force sensor to perform the palpation. The estimated stiffness map is then dynamically overlaid onto the registered preoperative model of the organ. We demonstrate the efficacy of our system by performing experiments on a phantom prostate model and other silicone organs with embedded stiff inclusions using the da Vinci research kit.

To be successful in multi-player drone racing, a player must not only follow the race track in an optimal way, but also compete with other drones through strategic blocking, faking, and opportunistic passing while avoiding collisions. Since unveiling one's own strategy to the adversaries is not desirable, this requires each player to independently predict the other players' future actions. Nash equilibria are a powerful tool to model this and similar multi-agent coordination problems in which the absence of communication impedes full coordination between the agents. In this paper, we propose a novel receding horizon planning algorithm that, exploiting sensitivity analysis within an iterated best response computational scheme, can approximate Nash equilibria in real time. We demonstrate that our solution effectively competes against alternative strategies in a large number of drone racing simulations.

We developed a novel whisker-follicle sensor that measures three mechanical signals at the whisker base. The first two signals are closely related to the two bending moments, and the third is an approximation to the axial force. Previous simulation studies have shown that these three signals are sufficient to determine the three-dimensional (3D) location at which the whisker makes contact with an object. Here we demonstrate hardware implementation of 3D contact point determination and then use continuous sweeps of the whisker to show proof-of principle 3D contour extraction. We begin by using simulations to confirm the uniqueness of the mapping between the mechanical signals at the whisker base and the 3D contact point location for the specific dimensions of the hardware whisker. Multi-output random forest regression is then used to predict the contact point locations of objects based on observed mechanical signals. When calibrated to the simulated data, signals from the hardware whisker can correctly predict contact point locations to within 1.5 cm about 74% of the time. However, if normalized output voltages from the hardware whiskers are used to train the algorithm (without calibrating to simulation), predictions improve to within 1.5 cm for about 96% of contact points and to within 0.6 cm for about 78% of contact points. This improvement suggests that as long as three appropriate predictor signals are chosen, calibrating to simulations may not be required. The sensor was next used to perform contour extraction on a cylinder and a cone. We show that basic contour extraction can be obtained with just two sweeps of the sensor. With further sweeps, it is expected that full 3D shape reconstruction could be achieved.