ICRA 2021 Papers

Industrial automation requires robot dexterity to automate many processes such as product assembling, packaging, and material handling. The existing robotic systems lack the capability to determining proper grasp strategies in the context of object affordances and task designations. In this paper, a framework of task-oriented dexterous grasping is proposed to learn grasp knowledge from human experience and to deploy the grasp strategies while adapting to grasp context. Grasp topology is defined and grasp strategies are learned from an established dataset for task-oriented dexterous manipulation. To adapt to various grasp context, a reinforcement-learning based grasping policy was implemented to deploy different task-oriented strategies. The performance of the system was evaluated in a simulated grasping environment by using an AR10 anthropomorphic hand installed in a Sawyer robotic arm. The proposed framework achieved a hit rate of 100% for grasp strategies and an overall top-3 match rate of 95.6%. The success rate of grasping was 85.6% during 2700 grasping experiments for manipulation tasks given in natural-language instructions.

Recent work has demonstrated the success of reinforcement learning (RL) for training bipedal locomotion policies for real robots. This prior work, however, has focused on learning joint-coordination controllers based on an objective of following joint trajectories produced by already available controllers. As such, it is difficult to train these approaches to achieve higher-level goals of legged locomotion, such as simply specifying the desired end-effector foot movement or ground reaction forces. In this work, we propose an approach for integrating knowledge of the robot system into RL to allow for learning at the level of task space actions in terms of feet setpoints. In particular, we integrate learning a task space policy with a model-based inverse dynamics controller, which translates task space actions into joint-level controls. With this natural action space for learning locomotion, the approach is more sample efficient and produces desired task space dynamics compared to learning purely joint space actions. We demonstrate the approach in simulation and also show that the learned policies are able to transfer to the real bipedal robot Cassie. This result encourages further research towards incorporating bipedal control techniques into the structure of the learning process to enable dynamic behaviors.

We’re interested in the problem of estimating object states from touch during manipulation under occlusions. In this work, we address the problem of estimating object poses from touch during planar pushing. Vision-based tactile sensors provide rich, local image measurements at the point of contact. A single such measurement, however, contains limited information and multiple measurements are needed to infer latent object state. We solve this inference problem using a factor graph. In order to incorporate tactile measurements in the graph, we need local observation models that can map highdimensional tactile images onto a low-dimensional state space. Prior work has used low-dimensional force measurements or engineered functions to interpret tactile measurements. These methods, however, can be brittle and difficult to scale across objects and sensors. Our key insight is to directly learn tactile observation models that predict the relative pose of the sensor given a pair of tactile images. These relative poses can then be incorporated as factors within a factor graph. We propose a two-stage approach: first we learn local tactile observation models supervised with ground truth data, and then integrate these models along with physics and geometric factors within a factor graph optimizer. We demonstrate reliable object tracking using only tactile feedback for ~150 real-world planar pushing sequences with varying trajectories across three object shapes.

Existing studies demonstrated that surgical motion patterns are strongly correlated with surgical outcomes. Real surgeries are complicated and it is expensive to harvest surgical data. Consequently, existing researches on surgical motion patterns focus on specific concise surgical tasks or simple surgical procedures. The paper presents a surgical motion pattern modeling technique that uses small data but can be applied to virtually any Endoscopic Sinus and Skull Base Surgeries (ESSBSs). The proposed method decreases the dimensionalities of the feature space through projecting surgical instrument motions into the endoscope coordinate, based on human expert domain knowledge. Furthermore, the method uses kinematic features and learns the motion pattern with Gaussian Process learning techniques. Comparing with existing surgical motion pattern modeling methods, the proposed method: 1, learns the motion model from small data; 2, can be generally applied to ESSBSs because it neither assumes nor depends on specific surgical tasks; 3, provides informative results in a real-time manner for optimizing surgical motions for improving surgical outcomes. The proposed method was verified by predicting surgical skill levels on cadaver surgeries. The results show the real-time prediction precision is higher than 81% and the offline accumulated precision reach 100%.

Reinforcement Learning (RL) of robotic manipulation skills, despite its impressive successes, stands to benefit from incorporating domain knowledge from control theory. One of the most important properties that is of interest is control stability. Ideally, one would like to achieve stability guarantees while staying within the framework of state-of-the-art deep RL algorithms. Such a solution does not exist in general, especially one that scales to complex manipulation tasks. We contribute towards closing this gap by introducing normalizing-flow control structure, that can be deployed in any latest deep RL algorithms. While stable exploration is not guaranteed, our method is designed to ultimately produce deterministic controllers with provable stability. In addition to demonstrating our method on challenging contact-rich manipulation tasks, we also show that it is possible to achieve considerable exploration efficiency–reduced state space coverage and actuation efforts– without losing learning efficiency.

Bottom-up approaches for image-based multi-person pose estimation consist of two stages: (1) keypoint detection and (2) grouping of the detected keypoints to form person instances. Current grouping approaches rely on learned embedding from only visual features that completely ignore the spatial configuration of human poses. In this work, we formulate the grouping task as a graph partitioning problem, where we learn the affinity matrix with a Graph Neural Network (GNN). More specifically, we design a Geometry-aware Association GNN that utilizes spatial information of the keypoints and learns local affinity from the global context. The learned geometry-based affinity is further fused with appearance-based affinity to achieve robust keypoint association. Spectral clustering is used to partition the graph for the formation of the pose instances. Experimental results on two benchmark datasets show that our proposed method outperforms existing appearance-only grouping frameworks, which shows the effectiveness of utilizing spatial context for robust grouping. Source code is available at: https://github.com/jiahaoLjh/PoseGrouping.

Deformable object manipulation tasks have long been regarded as challenging robotic problems. However, until recently very little work has been done on the subject, with most robotic manipulation methods being developed for rigid objects. Deformable objects are more difficult to model and simulate, which has limited the use of model-free Reinforcement Learning (RL) strategies, due to their need for large amounts of data that can only be satisfied in simulation. This paper proposes a new shape control task for Deformable Linear Objects (DLOs). More notably, we present the first study on the effects of elastoplastic properties on this type of problem. Objects with elastoplasticity such as metal wires, are found in various applications and are challenging to manipulate due to their nonlinear behavior. We first highlight the challenges of solving such a manipulation task from an RL perspective, particularly in defining the reward. Then, based on concepts from differential geometry, we propose an intrinsic shape representation using discrete curvature and torsion. Finally, we show through an empirical study that in order to successfully solve the proposed task using Deep Deterministic Policy Gradient (DDPG), the reward needs to include intrinsic information about the shape of the DLO.

This paper explores the idea that skillful assembly is best represented as dynamic sequences of Manipulation Primitives, and that such sequences can be automatically discovered by Reinforcement Learning. Manipulation Primitives, such as "Move down until contact", "Slide along x while maintaining contact with the surface", have enough complexity to keep the search tree shallow, yet are generic enough to generalize across a wide range of assembly tasks. Moreover, the additional "semantics" of the Manipulation Primitives make them more robust in sim2real and against model/environment variations and uncertainties, as compared to more elementary actions. Policies are learned in simulation, and then transferred onto a physical platform. Direct sim2real transfer (without retraining in real) achieves excellent success rates on challenging assembly tasks, such as round peg insertion with 0.04mm clearance or square peg insertion with large hole position/orientation estimation errors.

Polyculture farming is a sustainable farming technique based on synergistic interactions between differing plant types that make them more resistant to diseases and pests and better able to retain water. Reduced uniformity can reduce use of pesticides, fertilizer, and water, but is more labor intensive and more challenging to automate. We describe a scaled physical testbed (1.5m×3.0m) that uses a high resolution camera and soil sensors to monitor polyculture plants to facilitate tuning of plant growth, companion effects, and irrigation parameters for a first-order garden simulator. We use this simulator to develop a novel seed placement algorithm that increases coverage and diversity, and a learned pruning policy. In simulation experiments, the seed placement algorithm yields 60% more coverage and 10% more diversity than random seed placement and the learned pruning policy runs 1000X faster than a procedural lookahead policy to achieve high leaf coverage and plant diversity on adversarial gardens that include plant species with diverse growth rates. These models and policies provide the groundwork for a fully-automated system under development. Code, datasets and supplementary material can be found at https://github.com/BerkeleyAutomation/AlphaGarden/.

Earlier work has shown that reusing experience from prior motion planning problems can improve the efficiency of similar, future motion planning queries. However, for robots with many degrees-of-freedom, these methods exhibit poor generalization across different environments and often require large datasets that are impractical to gather. We present SPARK and FLAME, two experience-based frameworks for sampling-based planning applicable to complex manipulators in 3D environments. Both combine samplers associated with features from a workspace decomposition into a global biased sampling distribution. SPARK decomposes the environment based on exact geometry while FLAME is more general, and uses an octree-based decomposition obtained from sensor data. We demonstrate the effectiveness of SPARK and FLAME on a real and simulated Fetch robot tasked with challenging pick-and-place manipulation problems. Our approaches can be trained incrementally and significantly improve performance with only a handful of examples, generalizing better over diverse tasks and environments as compared to prior approaches.

We present an approach to learn fast and dynamic robot motions without exceeding limits on the position θ, velocity \dot \theta \dot \theta , acceleration \ddot \theta \ddot \theta and jerk \dddot \theta \dddot \theta of each robot joint. Movements are generated by mapping the predictions of a neural network to safely executable joint accelerations. The neural network is invoked periodically and trained via reinforcement learning. Our main contribution is an analytical procedure for calculating safe joint accelerations, which considers the prediction frequency fN of the neural network. As a result, the frequency fN can be freely chosen and treated as a hyperparameter. We show that our approach is preferable to penalizing constraint violations as it provides explicit guarantees and does not distort the desired optimization target. In addition, the influence of the selected prediction frequency on the learning performance and on the computing effort is highlighted by various experiments.

Humans leverage the dynamics of the environment and their own bodies to accomplish challenging tasks such as grasping an object while walking past it or pushing off a wall to turn a corner. Such tasks often involve switching dynamics as the robot makes and breaks contact. Learning these dynamics is a challenging problem and prone to model inaccuracies, especially near contact regions. In this work, we present a framework for learning composite dynamical behaviors from expert demonstrations. We learn a switching linear dynamical model with contacts encoded in switching conditions as a close approximation of our system dynamics. We then use discrete-time LQR as the differentiable policy class for data-efficient learning of control to develop a control strategy that operates over multiple dynamical modes and takes into account discontinuities due to contact. In addition to predicting interactions with the environment, our policy effectively reacts to inaccurate predictions such as unanticipated contacts. Through simulation and real world experiments, we demonstrate generalization of learned behaviors to different scenarios and robustness to model inaccuracies during execution.

Validating the kinematic feasibility of a planned robot motion and finding corresponding inverse solutions are time-consuming processes, especially for long-horizon manipulation tasks. Most existing approaches are based on solving iterative gradient-based optimization, so the processes are time-consuming and have a high risk of falling in local minima. In this work, we propose a unified framework to learn a kinematic feasibility model and a one-shot inverse mapping model for a redundant robot manipulator. Once they are trained, the models can compute the kinematic reachability of a target pose and its inverse solutions without iterative process. We validate our approach using a 7-DOF robot arm with an object grasping application.

Panoptic Segmentation aims to provide an understanding of background (stuff) and instances of objects (things) at a pixel level. It combines the separate tasks of semantic segmentation (pixel level classification) and instance segmentation to build a single unified scene understanding task. Typically, panoptic segmentation is derived by combining semantic and instance segmentation tasks that are learned separately or jointly (multi-task networks). In general, instance segmentation networks are built by adding a foreground mask estimation layer on top of object detectors or using instance clustering methods that assign a pixel to an instance center. In this work, we present a fully convolution neural network that learns instance segmentation from semantic segmentation and instance contours (boundaries of things). Instance contours along with semantic segmentation yield a boundary aware semantic segmentation of things. Connected component labeling on these results produces instance segmentation. We merge semantic and instance segmentation results to output panoptic segmentation. We evaluate our proposed method on the CityScapes dataset to demonstrate qualitative and quantitative performances along with several ablation studies. Our overview video can be accessed from https://youtu.be/wBtcxRhG3e0.

Addressing on monocular visual odometry problem, this paper presents a novel end-to-end network for estimation of camera ego-motion. The network learns the latent space of optical flow (OF) and models sequential dynamics so that the motion estimation is constrained by the relations between sequential images. We compute the OF field of consecutive images and extract the latent OF representation in a self-encoding manner. A Recurrent Neural Network is then followed to examine the OF changes, i.e., to conduct sequential learning. The extracted sequential OF latent space is used to compute the regression of the 6-dimensional pose vector. Particularly, we separately train the encoder in an unsupervised manner. By this means, we avoid non-convergence during the training of the whole network and allow more generalized and effective feature representation. Substantial experiments have been conducted on KITTI and Malaga datasets, and the results demonstrate that our model outperforms most learning-based VO approaches.

Everyday contact-rich tasks, such as peeling, cleaning, and writing, demand multimodal perception for effective and precise task execution. However, these present a novel challenge to robots as they lack the ability to combine these multimodal stimuli for performing contact-rich tasks. Learning-based methods have attempted to model multi-modal contact-rich tasks, but they often require extensive training examples and task-specific reward functions which limits their practicality and scope. Hence, we propose a generalizable model-free learning-from-demonstration framework for robots to learn contact-rich skills without explicit reward engineering. We present a novel multi-modal sensor data representation which improves the learning performance for contact-rich skills. We performed training and experiments using the real-life Sawyer robot for three everyday contact-rich skills – cleaning, writing, and peeling. Notably, the framework achieves a success rate of 100% for the peeling and writing skill, and 80% for the cleaning skill. Hence, this skill learning framework can be extended for learning other physical manipulation skills.

Imitation Learning (IL) is a powerful paradigm to teach robots to perform manipulation tasks by allowing them to learn from human demonstrations collected via teleoperation, but has mostly been limited to single-arm manipulation. However, many real-world tasks require multiple arms, such as lifting a heavy object or assembling a desk. Unfortunately, applying IL to multi-arm manipulation tasks has been challenging –asking a human to control more than one robotic arm can impose significant cognitive burden and is often only possible for a maximum of two robot arms. To address these challenges, we present MULTI-ARM ROBOTURK (MART), a multi-user data collection platform that allows multiple remote users to simultaneously teleoperate a set of robotic arms and collect demonstrations for multi-arm tasks. Using MART, we collected demonstrations for five novel two and three-arm tasks from several geographically separated users. From our data we arrived at a critical insight: most multi-arm tasks do not require global coordination throughout its full duration, but only during specific moments. We show that learning from such data consequently presents challenges for centralized agents that directly attempt to model all robot actions simultaneously, and perform a comprehensive study of different policy architectures with varying levels of centralization on our tasks. Finally, we propose and evaluate a base-residual policy framework that allows trained policies to better adapt to the mixed coordination setting common in multi-arm manipulation, and show that a centralized policy augmented with a decentralized residual model outperforms all other models on our set of benchmark tasks. Additional results and videos at https://roboturk.stanford.edu/multiarm

Human interaction involves very sophisticated non-verbal communication skills like understanding the goals and actions of others and coordinating our own actions accordingly. Neuroscience refers to this mechanism as motor resonance, in the sense that the perception of another persons actions and sensory experiences activates the observer’s brain as if (s)he would be performing the same actions and having the same experiences.We analyze and model the non-verbal cues exchanged between two humans in handover actions. The contributions of this paper are the following: (i) computational models, using recorded motion data, describing the motor behaviour of each actor in action-in-interaction situations; (ii) a computational model that captures the behaviour of the "giver" and "receiver" during an object handover action, by coupling the wrist kinematic motion of the actors; and (iii) the transfer of these models to the iCub robot for both action execution and recognition.Our results show that: (i) the robot is able to interpret the human wrist motion and infer whether or not the observed action is an "handover"; and (ii) use the motor resonance model to coordinate its actions with the human partner, during handover actions.

Recently, deep-learning based approaches have achieved impressive performance for autonomous driving. However, end-to-end vision-based methods typically have limited interpretability, making the behaviors of the deep networks difficult to explain. Hence, their potential applications could be limited in practice. To address this problem, we propose an interpretable end-to-end vision-based motion planning approach for autonomous driving, referred to as IVMP. Given a set of past surrounding-view images, our IVMP first predicts future egocentric semantic maps in bird’s-eye-view space, which are then employed to plan trajectories for self-driving vehicles. The predicted future semantic maps not only provide useful interpretable information, but also allow our motion planning module to handle objects with low probability, thus improving the safety of autonomous driving. Moreover, we also develop an optical flow distillation paradigm, which can effectively enhance the network while still maintaining its real-time performance. Extensive experiments on the nuScenes dataset and closed-loop simulation show that our IVMP significantly outperforms the state-of-the-art approaches in imitating human drivers with a much higher success rate. Our project page is available at https://sites.google.com/view/ivmp.