RSS 2019 Papers

State-of-the-art sampling-based online POMDP solvers compute near-optimal policies for POMDPs with very large state spaces. However, when faced with large observation spaces, they may become overly optimistic and compute sub-optimal policies, because of particle divergence. This paper presents a new online POMDP solver DESPOT alpha, which builds upon the widely used DESPOT solver. DESPOT alpha improves the practical performance of online planning for POMDPs with large observation as well as state spaces. Like DESPOT, DESPOT alpha uses the particle belief approximation and searches a determinized sparse belief tree. To tackle large observation spaces, DESPOT alpha shares sub-policies among many observations during online policy computation. The value function of a sub-policy is a linear function of the belief, commonly known as alpha-vector. We introduce a particle approximation of the alpha-vector to improve the efficiency of online policy search. We further speed up DESPOT alpha using CPU and GPU parallelization ideas introduced in HyP-DESPOT. Experimental results show that DESPOT Alpha/HyP-DESPOT Alpha; outperform DESPOT/HyP-DESPOT on POMDPs with large observation spaces, including a complex simulation task involving an autonomous vehicle driving among many pedestrians.

Robots that need to act in an uncertain, populated, and varied world need heterogeneous sensors to be able to perceive and act robustly. For example, self-driving cars currently on the road are equipped with dozens of sensors of several types (lidar, radar, sonar, cameras, ...). All of this existing and emerging complexity opens up many interesting questions regarding how to deal with multi-modal perception and learning. The recently developed technique of "wormhole learning"; shows that even temporary access to a different sensor with complementary invariance characteristics can be used to enlarge the operating domain of an existing object detector without the use of additional training data. For example, an RGB object detector trained with daytime data can be updated to function at night time by using a "wormhole" jump through a different modality that is more illumination invariant, such as an IR camera. It turns out that having an additional sensor improves performance, even if you subsequently lose it. In this work we extend wormhole learning to allow to cope with sensors that are radically different, such as RGB cameras and event-based neuromorphic sensors. Their profound differences imply that we need a more careful selection of which samples to transfer, by the definition of "cross-modal learning filters". We will walk in a relatively unexplored territory of multi-modal observability that is not usually considered in machine learning. We show that wormhole learning increases performance even though the intermediate neuromorphic modality is on average much worse at the task. These results suggest that multi-modal learning for perception is still an early field and there might be many opportunities to improve the perception performance by accessing a rich set of heterogeneous sensors (even if some are not actually deployed on the robot).

This paper reports on a novel formulation and evaluation of visual odometry from RGB-D images. Assuming a static scene, the developed theoretical framework generalizes the widely used direct energy formulation (photometric error minimization) technique for obtaining a rigid body transformation that aligns two overlapping RGB-D images to a continuous formulation. The continuity is achieved through functional treatment of the problem and representing the process models over RGB-D images in a reproducing kernel Hilbert space; consequently, the registration is not limited to the specific image resolution and the framework is fully analytical with a closed-form derivation of the gradient. We solve the problem by maximizing the inner product between two functions defined over RGB-D images, while the continuous action of the rigid body motion Lie group is captured through the integration of the flow in the corresponding Lie algebra. Energy-based approaches have been extremely successful and the developed framework in this paper shares many of their desired properties such as the parallel structure on both CPUs and GPUs, sparsity, semi-dense tracking, avoiding explicit data association which is computationally expensive, and possible extensions to the simultaneous localization and mapping frameworks. The evaluations on experimental data and comparison with the equivalent energy-based formulation of the problem confirm the effectiveness of the proposed technique, especially, when the lack of structure and texture in the environment is evident.

Conditional Neural Movement Primitives (CNMPs) is a learning from demonstration framework that is designed as a robotic movement learning and generation system built on top of a recent deep neural architecture, namely Conditional Neural Processes (CNPs). Based on CNPs, CNMPs extract the prior knowledge directly from the training data by sampling observations from it, and uses it to predict a conditional distribution over any other target points. CNMPs specifically learns complex temporal multi-modal sensorimotor relations in connection with external parameters and goals; produces movement trajectories in joint or task space; and executes these trajectories through a high-level feedback control loop. Conditioned with an external goal that is encoded in the sensorimotor space of the robot, predicted sensorimotor trajectory that is expected to be observed during the successful execution of the task is generated by the CNMP, and the corresponding motor commands are executed. In order to detect and react to unexpected events during action execution, CNMP is further conditioned with the actual sensor readings in each time-step. Through simulations and real robot experiments, we showed that CNMPs can learn the non-linear relations between low-dimensional parameter spaces and complex movement trajectories from few demonstrations; and they can also model the associations between high-dimensional sensorimotor spaces and complex motions using large number of demonstrations. The experiments further showed that even the task parameters were not explicitly provided to the system, the robot could learn their influence by associating the learned sensorimotor representations with the movement trajectories. The robot, for example, learned the influence of object weights and shapes through exploiting its sensorimotor space that includes proprioception and force measurements; and be able to change the movement trajectory on the fly when one of these factors were changed through external intervention.

Algorithms based on deep network models are being used for many pattern recognition and decision-making tasks in robotics and AI. Training these models requires a large labeled dataset and considerable computational resources, which are not readily available in many domains. Also, it is difficult to understand the internal representations and reasoning mechanisms of these models. The architecture described in this paper attempts to address these limitations by drawing inspiration from research in cognitive systems. It uses non-monotonic logical reasoning with incomplete commonsense domain knowledge, and inductive learning of previously unknown constraints on the domain's states, to guide the construction of deep network models based on a small number of relevant training examples. As a motivating example, we consider a robot reasoning about the stability and partial occlusion of configurations of objects in simulated images. Experimental results indicate that in comparison with an architecture based just on deep networks, our architecture improves reliability, and reduces the sample complexity and time complexity of training deep networks.

Untethered magnetically actuated microrobots can access distant, enclosed and small spaces, such as inside microfluidic channels and the human body, making them appealing for minimal invasive tasks. Despite the simplicity of individual magnetic microrobots, a collective of these microrobots that can work closely and cooperatively would significantly enhance their capabilities. However, a challenge of realizing such collective magnetic microrobots is to coordinate their formations and motions with under-actuated control signals. Here, we report a method that allows collective magnetic microrobots working closely and cooperatively by controlling their two-dimensional (2D) formations and collective motions in a programmable and reconfigurable manner. The actively designed formation and intrinsic adjustable compliance within the group allow bio-inspired collective behaviors, such as navigating through cluttered environments and reconfigurable cooperative manipulation ability. These collective magnetic microrobots thus could enable potential applications in programmable self-assembly, modular robotics, swarm systems, and biomedicine.

Our goal is to train a policy for autonomous driving via imitation learning that is robust enough to drive a real vehicle. We find that standard behavior cloning is insufficient for handling complex driving scenarios, even when we leverage a perception system for preprocessing the input and a controller for executing the output on the car: 30 million examples are still not enough. We propose exposing the learner to synthesized data in the form of perturbations to the expert's driving, which creates interesting situations such as collisions and/or going off the road. Rather than purely imitating all data, we augment the imitation loss with additional losses that penalize undesirable events and encourage progress -- the perturbations then provide an important signal for these losses and lead to robustness of the learned model. We show that the ChauffeurNet model can handle complex situations in simulation, and present ablation experiments that emphasize the importance of each of our proposed changes and show that the model is responding to the appropriate causal factors. Finally, we demonstrate the model driving a real car at our test facility.

The problem of temporal alignment of time series is common across many fields of study. Within the domain of robotics, human motion trajectories are one type of time series that is often utilized for recognition and prediction of human intent. In these applications, online temporal alignment of partial trajectories to a full representative trajectory is of particular interest, as it is desirable to make accurate intent prediction decisions early in a motion in order to enable proactive robot behavior. This is a particularly difficult problem, however, due to the potential for overlapping trajectory regions and temporary stops, both of which can degrade the performance of existing alignment techniques. Furthermore, it is desirable to not only provide the most likely alignment but also characterize the uncertainty around it, which current methods are unable to accomplish. To address these difficulties and drawbacks, we present BEST-PTA, a framework that combines optimization, supervised learning, and unsupervised learning components in order to build a Bayesian model that outputs distributions over likely correspondence points based on observed partial trajectory data. Through an evaluation incorporating multiple datasets, we show that BEST-PTA outperforms previous alignment techniques; furthermore, we demonstrate that this improvement can significantly boost human motion prediction performance and discuss the implications of these results on improving the quality of human-robot interaction.

We introduce BayesSim, a framework for robotics simulations allowing a full Bayesian treatment for the parameters of the simulator. As simulators become more sophisticated and able to represent the dynamics more accurately, fundamental problems in robotics such as motion planning and perception can be solved in simulation and solutions transferred to the physical robot. However, even the most complex simulator might still not be able to represent reality in all its details either due to inaccurate parametrization or simplistic assumptions in the dynamic models. BayesSim provides a principled framework to reason about the uncertainty of simulation parameters. Given a black-box simulator (or generative model) that outputs trajectories of state and action pairs from unknown simulation parameters, followed by trajectories obtained with a physical robot, we develop a likelihood-free inference method that computes the posterior distribution of simulation parameters. This posterior can then be used in problems where Sim2Real is critical, for example in policy search. We compare the performance of BayesSim in obtaining accurate posteriors in a number of classical control and robotics problems, and show that the posterior computed from BayesSim can be used for domain radomization outperforming alternative methods that randomize based on uniform priors.

This work explores the problem of robot tool construction - creating tools from parts available in the environment. We advance the state-of-the-art in robotic tool construction by introducing an approach that enables the robot to construct a wider range of tools with greater computational efficiency. Specifically, given an action that the robot wishes to accomplish and a set of building parts available to the robot, our approach reasons about the shape of the parts and potential ways of attaching them, generating a ranking of part combinations that the robot then uses to construct and test the target tool. We validate our approach on the construction of five tools using a physical 7-DOF robot arm.

A robot's mechanical parts routinely wear out from normal functioning and can be lost to injury. For autonomous robots operating in isolated or hostile environments, repair from a human operator is often not possible. Thus, much work has sought to automate damage recovery in robots. However, every case reported in the literature to date has accepted the damaged mechanical structure as fixed, and focused on learning new ways to control it. Here we show for the first time a robot that automatically recovers from unexpected damage by deforming its resting mechanical structure without changing its control policy. We found that, especially in the case of "deep insult" such as removal of all four of the robot's legs, the damaged machine evolves shape changes that not only recover the original level of function (locomotion) as before, but can in fact surpass the original level of performance (speed). This suggests that shape change, instead of control readaptation, may be a better method to recover function after damage in some cases.

This paper proposes a novel highly scalable sampling-based planning algorithm for multi-robot active information acquisition tasks in complex environments. Active information gathering scenarios include target localization and tracking, active SLAM, surveillance, environmental monitoring and others. The objective is to compute control policies for sensing robots which minimize the accumulated uncertainty of a dynamic hidden state over an a priori unknown horizon. To address this problem, we propose a new sampling-based algorithm that simultaneously explores both the robot motion space and the reachable information space. Unlike relevant sampling-based approaches, we show that the proposed algorithm is probabilistically complete, asymptotically optimal and is supported by convergence rate bounds. Moreover, we demonstrate that by introducing bias in the sampling process towards informative areas, the proposed method can quickly compute sensor policies that achieve desired levels of uncertainty in large-scale estimation tasks that may involve large sensor teams, workspaces, and dimensions of the hidden state. We provide extensive simulation results that corroborate the theoretical analysis and show that the proposed algorithm can address large-scale estimation tasks which were previously infeasible.

Model-based approaches bear great promise for decision making of agents interacting with the physical world. In the context of spatial environments, different types of problems such as localisation, mapping, navigation or autonomous exploration are typically adressed with specialised methods, often relying on detailed knowledge of the system at hand. We express these tasks as probabilistic inference and planning under the umbrella of deep sequential generative models. Using the frameworks of variational inference and neural networks, our method inherits favourable properties such as flexibility, scalability and the ability to learn from data. The method performs comparably to specialised state-of-the-art methodology in two distinct simulated environments.

Model predictive control (MPC) is a powerful technique for solving dynamic control tasks. In this paper, we show that there exists a close connection between MPC and online learning, an abstract theoretical framework for analyzing online decision making in the optimization literature. This new perspective provides a foundation for leveraging powerful online learning algorithms to design MPC algorithms. Specifically, we propose a new algorithm based on dynamic mirror descent (DMD), an online learning algorithm that is designed for non-stationary setups. Our algorithm, Dynamic Mirror Descent Model Predictive Control (DMD-MPC), represents a general family of MPC algorithms that includes many existing techniques as special instances. DMD-MPC also provides a fresh perspective on previous heuristics used in MPC and suggests a principled way to design new MPC algorithms. In the experimental section of this paper, we demonstrate the flexibility of DMD-MPC, presenting a set of new MPC algorithms on a simple simulated cartpole and a simulated and real-world aggressive driving task. A video of the real-world experiment can be found at https://youtu.be/vZST3v0_S9w.

This paper presents an omnidirectional aerial manipulation platform for robust and responsive interaction with unstructured environments, toward the goal of contact-based inspection. The fully actuated tilt-rotor aerial system is equipped with a rigidly mounted end-effector, and is able to exert a 6 degree of freedom force and torque, decoupling the system's translational and rotational dynamics, and enabling precise interaction with the environment while maintaining stability. An impedance controller with selective apparent inertia is formulated to permit compliance in certain degrees of freedom while achieving precise trajectory tracking and disturbance rejection in others. Experiments demonstrate disturbance rejection, push-and-slide interaction, and on-board state estimation with depth servoing to interact with local surfaces. The system is also validated as a tool for contact-based non-destructive testing of concrete infrastructure.

We propose a robust approach for the registration of two sets of 3D points in the presence of a large amount of outliers. Our first contribution is to reformulate the registration problem using a Truncated Least Squares (TLS) cost that makes the estimation insensitive to a large fraction of spurious point- to-point correspondences. The second contribution is a general framework to decouple rotation, translation, and scale estimation, which allows solving in cascade for the three transformations. Since each subproblem (scale, rotation, and translation estimation) is still non-convex and combinatorial in nature, out third contribution is to show that (i) TLS scale and (component-wise) translation estimation can be solved exactly and in polynomial time via an adaptive voting scheme, (ii) TLS rotation estimation can be relaxed to a semidefinite program and the relaxation is tight in practice, even in the presence of an extreme amount of outliers. We validate the proposed algorithm, named TEASER (Truncated least squares Estimation And SEmidefinite Relaxation), in standard registration benchmarks showing that the algorithm outperforms RANSAC and robust local optimization techniques, and favorably compares with Branch-and-Bound methods, while being a polynomial-time algorithm. TEASER can tolerate up to 99% outliers and returns highly-accurate solutions.

This work approaches the challenge of how to divide the problem of Simultaneous Localization and Mapping (SLAM) into its smallest possible constituents, in such a way that the reusability and interchangeability of each such module is maximized. In particular, most components in the proposed system should be not aware of details such that whether the map comprises a single global map or a set of local submaps, whether the state vector is defined in SE(2) or SE(3), with or without velocity, etc. Any number of heterogeneous sensors should be used together and their information fused seamlessly into a consistent localization solution. The resulting system would be useful for researchers, easing the development of reproducible research and enabling the quick adoption of state-of-the-art algorithms into product prototypes. Our implementation has been tested with different sensors against the KITTI, EuRoC, and KAIST datasets. In this paper we focus on an introduction to the framework and on experimental results for 3D LiDAR odometry and mapping. LiDAR SLAM for the KITTI datasets achieves typical translation errors between 0.5% and 2% for most urban sequences, while processing the data at 1.5x the real-time rate with a reduced memory requirement thanks to our framework's capability to dynamically swap out from memory the parts of the map that are not immediately required, transparently loading them again when required. The framework will be released as open-source at https://github.com/MOLAorg/mola

Untethered small-scale soft robots can potentially be used in healthcare and biomedical applications. They can access small spaces and reshape their bodies in a programmable manner to adapt to unstructured environments and have diverse dynamic behaviors. However, the functionalities of current miniature soft robots are limited, restricting their applications in medical procedures. Taking the advantage of the shape-programmable ability of magnetic soft composite materials, here we propose an untethered soft millirobot (jellyfishbot) that can swim like a jellyfish by time- and trajectory-asymmetric up and down beating of its lappets. Its swimming speed and direction can be controlled by tuning the magnitude, frequency, and direction of the external oscillating magnetic field. We demonstrate that such jellyfishbot can perform several tasks that could be useful towards medical applications, such as delivering drugs, clogging a narrow tube or vessel, and patching a target area under ultrasound imaging-based guiding. The millirobot presented in this paper could be used inside organs filled with fluids completely, such as a bladder or inflated stomach.

Motion planning for mobile robots with many degrees-of-freedom (DoF) is challenging due to their high-dimensional configuration spaces. To manage this curse of dimensionality, this paper proposes a new hierarchical framework that decomposes the system into sub-systems (based on shared capabilities of DoFs), for which we can design and coordinate motions. Instead of constructing a high-dimensional configuration space, we establish a hierarchy of two-dimensional spaces on which we can visually design gaits using geometric mechanics tools. We then coordinate motions among the two-dimensional spaces in a pairwise fashion to obtain desired robot locomotion. Further geometric analysis of the two-dimensional spaces allows us to visualize the contribution of each sub-system to the locomotion, as well as the contribution of the coordination among the sub-systems. We demonstrate our approach by designing gaits for quadrupedal robots with different morphologies, and experimentally validate our findings on a robot with a long actuated back and intermediate-sized legs.