ICRA 2018 Papers

Modeling biomechanical musculoskeletal systems reveals that the mapping from muscle stimulations to movement dynamics is highly nonlinear and complex, which makes it difficult to control those systems with classical techniques. In this work, we not only investigate whether machine learning approaches are capable of learning a controller for such systems. We are especially interested in the question if the structure of the musculoskeletal apparatus exhibits properties that are favorable for the learning task. In particular, we consider learning a control policy from target positions to muscle stimulations. To account for the high actuator redundancy of biomechanical systems, our approach uses a learned forward model represented by a neural network and sequential quadratic programming to obtain the control policy, which also enables us to alternate the co-contraction level and hence allows to change the stiffness of the system and to include optimality criteria like small muscle stimulations. Experiments on both a simulated musculoskeletal model of a human arm and a real biomimetic muscle-driven robot show that our approach is able to learn an accurate controller despite high redundancy and nonlinearity, while retaining sample efficiency.

In this paper we propose a new approach for efficient programming of grinding and polishing operation. In the proposed system, the initial policy is performed by a skilled operator and recorded with a passive digitizer. The demonstrated policy comprises both position and force data. The optimal robot execution of the task is provided by applying a virtual mechanism approach, which models the polishing/grinding tool as a serial kinematic chain. By joining the robot and the virtual mechanism in an augmented system, additional degrees of freedom are obtained and redundancy resolution can be applied to optimize the demonstrated motion. Another benefit of the proposed approach is that the same policy can be transferred to different combination of robots and grinding/polishing tools without any modification of the captured motion. The proposed approach requires known contact point between the treated object and the polishing/grinding tool. We propose a novel approach for accurate estimation of this point using data obtained from the force-torque sensor. Finally, the demonstrated path is refined to compensate for inaccurate calibration and different dynamics of a robot and the human demonstrator using iterative learning controller. The proposed method was verified in a real industrial environment.

In this paper we develop an approach for learning user preferences for complex task specifications through human-robot interaction. We consider the problem of planning robot motion in a known environment, but where a user has specified additional spatial and temporal constraints on allowable robot motions. To illustrate the impact of the user's constraints on performance, we iteratively present users with alternative solutions on an interface. The user provides a ranking of alternate paths, and from this we learn about the importance of different constraints. This allows for an accessible method for specifying complex robot tasks. We present an algorithm that iteratively builds a set of constraints on the relative importance of each user constraint, and prove that with sufficient interaction, the algorithm determines a user-optimal path. We demonstrate the practical performance by simulating realistic material transport scenarios in industrial facilities.

We explore beyond existing work in learning from demonstration by asking the question: “Can robots learn to guide?”, that is, can a robot autonomously learn an instructional policy from expert demonstration and use it to instruct humans in executing complex task? As a solution, we propose learning of instructional policy (πI) that maps the state to an instruction for a human. To learn πI, we define action primitives that addresses the challenge of mapping continuous state action trajectories to human parse-able instructions. Action primitives are demonstrated to be very effective in automatic segmentation of demonstration trajectories into fewer repetitive and reusable segments, and a highly scalable approach in comparison to the existing state-of-the art. Finally, we construct a non-generic policy model as a generative model for instructional policies to generate instruction for an entire task. With few modifications, the proposed model is demonstrated to perform autonomous execution of complex truck loading task on hydraulic actuated scaled excavator robot. Guidance approach is tested based on a controlled group study involving 75 participants, who learn to perform the same task.

Deep learning has been successfully applied to “end-to-end” learning of the autonomous driving task, where a deep neural network learns to predict steering control commands from camera data input. However, the learned representations do not support higher-level decision making required for autonomous navigation, nor the uncertainty estimates required for parallel autonomy, where vehicle control is shared between human and robot. This paper tackles the problem of learning a representation to predict a continuous control probability distribution, and thus steering control options and bounds for those options, which can be used for autonomous navigation. Each mode of the distribution encodes a possible macro-action that the system could execute at that instant, and the covariances of the modes place bounds on safe steering control values. Our approach has the added advantage of being trained on unlabeled data collected from inexpensive cameras. The deep neural network based algorithm generates a probability distribution over the space of steering angles, from which we leverage Variational Bayesian methods to extract a mixture model and compute the different possible actions in the environment. A bound, which the autonomous vehicle must respect in our parallel autonomy setting, is then computed for each of these actions. We evaluate our approach on a challenging dataset containing a wide variety of driving conditions, and show that our algorithm is capable of parameterizing Gaussian Mixture Models for possible actions, and extract steering bounds with a mean error of only 2 degrees. Additionally, we demonstrate our system working on a full scale autonomous vehicle and evaluate its ability to successful handle various different parallel autonomy situations.

In order to robustly execute a task under environmental uncertainty, a robot needs to be able to reactively adapt to changes arising in its environment. The environment changes are usually reflected in deviation from expected sensory traces. These deviations in sensory traces can be used to drive the motion adaptation, and for this purpose, a feedback model is required. The feedback model maps the deviations in sensory traces to the motion plan adaptation. In this paper, we develop a general data-driven framework for learning a feedback model from demonstrations. We utilize a variant of a radial basis function network structure -with movement phases as kernel centers- which can generally be applied to represent any feedback models for movement primitives. To demonstrate the effectiveness of our framework, we test it on the task of scraping on a tilt board. In this task, we are learning a reactive policy in the form of orientation adaptation, based on deviations of tactile sensor traces. As a proof of concept of our method, we provide evaluations on an anthropomorphic robot.

A defining feature of sampling-based motion planning is the reliance on an implicit representation of the state space, which is enabled by a set of probing samples. Traditionally, these samples are drawn either probabilistically or deterministically to uniformly cover the state space. Yet, the motion of many robotic systems is often restricted to “small” regions of the state space, due to e.g. differential constraints or collision-avoidance constraints. To accelerate the planning process, it is thus desirable to devise non-uniform sampling strategies that favor sampling in those regions where an optimal solution might lie. This paper proposes a methodology for nonuniform sampling, whereby a sampling distribution is learned from demonstrations, and then used to bias sampling. The sampling distribution is computed through a conditional variational autoencoder, allowing sample generation from the latent space conditioned on the specific planning problem. This methodology is general, can be used in combination with any sampling-based planner, and can effectively exploit the underlying structure of a planning problem while maintaining the theoretical guarantees of sampling-based approaches. Specifically, on several planning problems, the proposed methodology is shown to effectively learn representations for the relevant regions of the state space, resulting in an order of magnitude improvement in terms of success rate and convergence to the optimal cost.

Swarm robotics investigates how a large population of robots with simple actuation and limited sensors can collectively solve complex tasks. One particular interesting application with robot swarms is autonomous object assembly. Such tasks have been solved successfully with robot swarms that are controlled by a human operator using a light source. In this paper, we present a method to solve such assembly tasks autonomously based on policy search methods. We split the assembly process in two subtasks: generating a high-level assembly plan and learning a low-level object movement policy. The assembly policy plans the trajectories for each object and the object movement policy controls the trajectory execution. Learning the object movement policy is challenging as it depends on the complex state of the swarm which consists of an individual state for each agent. To approach this problem, we introduce a representation of the swarm which is based on Hilbert space embeddings of distributions. This representation is invariant to the number of agents in the swarm as well as to the allocation of an agent to its position in the swarm. These invariances make the learned policy robust to changes in the swarm and also reduce the search space for the policy search method significantly. We show that the resulting system is able to solve assembly tasks with varying object shapes in multiple simulation scenarios and evaluate the robustness of our representation to changes in the swarm size. Furthermore, we demonstrate that the policies learned in simulation are robust enough to be transferred to real robots.

In this work, motivated by recent manufacturing trends, we investigate autonomous robotic assembly. Industrial assembly tasks require contact-rich manipulation skills, which are challenging to acquire using classical control and motion planning approaches. Consequently, robot controllers for assembly domains are presently engineered to solve a particular task, and cannot easily handle variations in the product or environment. Reinforcement learning (RL) is a promising approach for autonomously acquiring robot skills that involve contact-rich dynamics. However, RL relies on random exploration for learning a control policy, which requires many robot executions, and often gets trapped in locally suboptimal solutions. Instead, we posit that prior knowledge, when available, can improve RL performance. We exploit the fact that in modern assembly domains, geometric information about the task is readily available via the CAD design files. We propose to leverage this prior knowledge by guiding RL along a geometric motion plan, calculated using the CAD data. We show that our approach effectively improves over traditional control approaches for tracking the motion plan, and can solve assembly tasks that require high precision, even without accurate state estimation. In addition, we propose a neural network architecture that can learn to track the motion plan, thereby generalizing the assembly controller to changes in the object positions.

Vision-based navigation is extremely susceptible to natural scene changes. This can result in localization failures in less than a few hours after map creation. To combat short-term illumination changes as well as long-term seasonal variations, we propose using a place-and-time-dependent binary descriptor that adapts to different scenarios in an online fashion. This is achieved by extending the GRIEF [6] evolution algorithm in two ways: correspondence generation using a known pose change and the inclusion of LATCH triplets in addition to BRIEF comparisons for descriptor generation. We show the adaptive descriptor outperforms a single descriptor scheme for localization within a single-experience Visual Teach and Repeat (VT&R) system while maintaining the efficiency of binary descriptors. By adapting the description function to different environmental conditions, it allows the system to operate for a longer period before a new experience is required. In the presence of extreme illumination changes from day to night, we obtain 40% more inlier matches compared to SURF. In the case of seasonal variations, a 70% increase is demonstrated. The increased correspondences result in more localizable sections along the paths, amounting to a 25% and 150% increase in the lighting and seasonal cases, respectively.

Constructing a smart wheelchair on a commercially available powered wheelchair (PWC) platform avoids a host of seating, mechanical design and reliability issues but requires methods of predicting and controlling the motion of a device never intended for robotics. Analog joystick inputs are subject to black-box transformations which may produce intuitive and adaptable motion control for human operators, but complicate robotic control approaches; furthermore, installation of standard axle mounted odometers on a commercial PWC is difficult. In this work, we present an integrated hardware and software system for predicting the motion of a commercial PWC platform that does not require any physical or electronic modification of the chair beyond plugging into an industry standard auxiliary input port. This system uses an RGB-D camera and an Arduino interface board to capture motion data, including visual odometry and joystick signals, via ROS communication. Future motion is predicted using an autoregressive sparse Gaussian process model. We evaluate the proposed system on real-world short-term path prediction experiments. Experimental results demonstrate the system's efficacy when compared to a baseline neural network model.

In this work, we investigate methods to detect four phenomena (modes) that occur during prehensile fingertip-based within-hand manipulation without the use of tactile sensors. By using actuator states and visual data, we aim to recognize different modes of operation such as interpreting if the hand is about to drop the object, if the object will begin to slide on the fingers, or if the system is at or near a singularity. For this purpose, we utilize supervised learning techniques, which allow us to detect the modes without the use of a mechanical model of the system. We analyze the individual roles of specific features available through both the actuator and visual data, and identify the ones that have the most significance for detecting the operation modes. Our results show classification performance of 96% (using either Extra Trees, Gradient Boosting, or SVM) when using combined actuator and visual features. Interestingly, we were able to achieve a 94% classification rate using only actuator information, and 93 % using only visual information. Overall, the classifiers identified actuator positions, actuator loads, and commanded velocities as the most important features for detecting a mode. These results have implications for enabling the control of within-hand manipulation movements utilizing a minimal amount of sensory information without a model of the hand/object system.

Complex contact manipulation tasks can be decomposed into sequences of motor primitives. Individual primitives often end with a distinct contact state, such as inserting a screwdriver tip into a screw head or loosening it through twisting. To achieve robust execution, the robot should be able to verify that the primitive's goal has been reached as well as disambiguate it from erroneous contact states. In this paper, we introduce and evaluate a framework to autonomously construct manipulation graphs from manipulation demonstrations. Our manipulation graphs include sequences of motor primitives for performing a manipulation task as well as corresponding contact state information. The sensory models for the contact states allow the robot to verify the goal of each motor primitive as well as detect erroneous contact changes. The proposed framework was experimentally evaluated on grasping, unscrewing, and insertion tasks on a Barrett arm and hand equipped with two BioTacs. The results of our experiments indicate that the learned manipulation graphs achieve more robust manipulation executions by confirming sensory goals as well as discovering and detecting novel failure modes.

This work presents an approach to learn path planning for robot social navigation by demonstration. We make use of Fully Convolutional Neural Networks (FCNs) to learn from expert's path demonstrations a map that marks a feasible path to the goal as a classification problem. The use of FCNs allows us to overcome the problem of manually designing/identifying the cost-map and relevant features for the task of robot navigation. The method makes use of optimal Rapidly-exploring Random Tree planner (RRT*) to overcome eventual errors in the path prediction; the FCNs prediction is used as cost-map and also to partially bias the sampling of the configuration space, leading the planner to behave similarly to the learned expert behavior. The approach is evaluated in experiments with real trajectories and compared with Inverse Reinforcement Learning algorithms that use RRT* as underlying planner.

Our goal is for people to be physically comfortable when taking objects from robots. This puts a burden on the robot to hand over the object in such a way that a person can easily reach it, without needing to strain or twist their arm - a way that is conducive to ergonomic human grasping configurations. To achieve this, the robot needs to understand what makes a configuration more or less ergonomic to the person, i.e. their ergonomic cost function. In this work, we formulate learning a person's ergonomic cost as an online estimation problem. The robot can implicitly make queries to the person by handing them objects in different configurations, and gets observations in response about the way they choose to take the object. We compare the performance of both passive and active approaches for solving this problem in simulation, as well as in an in-person user study.

Challenging tasks in unstructured environments require robots to learn complex models. Given a large amount of information, learning multiple simple models can offer an efficient alternative to a monolithic complex network. Training multiple models-that is, learning their parameters and their responsibilities-has been shown to be prohibitively hard as optimization is prone to local minima. To efficiently learn multiple models for different contexts, we thus develop a new algorithm based on expectation maximization (EM). In contrast to comparable concepts, this algorithm trains multiple modules of paired forward-inverse models by using the prediction errors of both forward and inverse models simultaneously. In particular, we show that our method yields a substantial improvement over only considering the errors of the forward models on tasks where the inverse space contains multiple solutions.

This paper focuses on the problem of learning 6- DOF grasping with a parallel jaw gripper in simulation. Our key idea is constraining and regularizing grasping interaction learning through 3D geometry prediction. We introduce a deep geometry-aware grasping network (DGGN) that decomposes the learning into two steps. First, we learn to build mental geometry-aware representation by reconstructing the scene (i.e., 3D occupancy grid) from RGBD input via generative 3D shape modeling. Second, we learn to predict grasping outcome with its internal geometry-aware representation. The learned outcome prediction model is used to sequentially propose grasping solutions via analysis-by-synthesis optimization. Our contributions are fourfold: (1) To best of our knowledge, we are presenting for the first time a method to learn a 6-DOF grasping net from RGBD input; (2) We build a grasping dataset from demonstrations in virtual reality with rich sensory and interaction annotations. This dataset includes 101 everyday objects spread across 7 categories, additionally, we propose a data augmentation strategy for effective learning; (3) We demonstrate that the learned geometry-aware representation leads to about 10% relative performance improvement over the baseline CNN on grasping objects from our dataset. (4) We further demonstrate that the model generalizes to novel viewpoints and object instances.

In this paper we consider the problem of autonomous exploration of an unknown, GPS-denied environment using a swarm of robots with very limited resources and limited sensing capabilities. To that end we use a landmark complex, a simplicial complex utilizing an observation of landmarks, as a topological representation of the environment. Each robot is equipped with an omni-directional, limited-range sensor that can identify landmarks in the robot's neighborhood. The robots use the bearing angles to the landmarks for local navigation. Given a collection of identifiable landmarks, a landmark complex can then be cumulatively constructed to encapsulate the topological information of the environment. Under the assumption of sufficiently dense landmarks, we propose an exploration and exploitation strategy that guides the swarm of robots to explore an environment using only bearing measurements without any metric information. Lastly, we demonstrate the coordinate-free and localization-free navigation in the environment using the constructed landmark complex.

Deep neural network (DNN) architectures have been shown to outperform traditional pipelines for object segmentation and pose estimation using RGBD data, but the performance of these DNN pipelines is directly tied to how representative the training data is of the true data. Hence a key requirement for employing these methods in practice is to have a large set of labeled data for your specific robotic manipulation task, a requirement that is not generally satisfied by existing datasets. In this paper we develop a pipeline to rapidly generate high quality RGBD data with pixelwise labels and object poses. We use an RGBD camera to collect video of a scene from multiple viewpoints and leverage existing reconstruction techniques to produce a 3D dense reconstruction. We label the 3D reconstruction using a human assisted ICP-fitting of object meshes. By reprojecting the results of labeling the 3D scene we can produce labels for each RGBD image of the scene. This pipeline enabled us to collect over 1,000,000 labeled object instances in just a few days. We use this dataset to answer questions related to how much training data is required, and of what quality the data must be, to achieve high performance from a DNN architecture. Our dataset and annotation pipeline are available at labelfusion.csail.mit.edu.