ICRA 2022 Papers

Existing navigation policies for autonomous robots tend to focus on collision avoidance while ignoring human-robot interactions in social life. For instance, robots can pass along the corridor safer and easier if pedestrians notice them. Sounds have been considered as an efficient way to attract the attention of pedestrians, which can alleviate the freezing robot problem. In this work, we present a new deep reinforcement learning (DRL) based social navigation approach for autonomous robots to move in pedestrian-rich environments with interaction capacity. Most existing DRL based methods intend to train a general policy that outputs both navigation actions, i.e., expected robot's linear and angular velocities, and interaction actions, i.e., the beep action, in the context of reinforcement learning. Different from these methods, we intend to train the policy via both supervised learning and reinforcement learning. In specific, we first train an interaction policy in the context of supervised learning, which provides a better understanding of the social situation, then we use this interaction policy to train the navigation policy via multiple reinforcement learning algorithms. We evaluate our approach in various simulation environments and compare it to other methods. The experimental results show that our approach outperforms others in terms of the success rate. We also deploy the trained policy on a real-world robot, which shows a nice performance in crowded environments.

When moving objects that are too bulky or heavy to be grasped or lifted, robotic manipulation can benefit from the object's interaction with the support surface and its natural dynamics under gravity. In this work, we show that such dynamic, underactuated manipulation capability can be acquired through reinforcement learning and deployed on real robot systems. First, we present a framework to learn a control policy for object transport in a dynamic simulation environment, featuring the object and the support surface. We then demonstrate successful object locomotion with the learned policy through a set of simulated and real-world experiments, performed with a robot arm and an aerial robot interacting with the object in a non-prehensile manner. While the object, which is in contact with the support surface, oscillates sideways passively under gravity, the robot uses the learned policy to move the object forward with a steady gait by regulating the mechanical energy and the posture of the object. Our experiment results show that the learned policy can transport the object through unmodeled effects of terrain and perturbation.

Recent work has demonstrated that motion planners' performance can be significantly improved by retrieving past experiences from a database. Typically, the experience database is queried for past similar problems using a similarity function defined over the motion planning problems. However, to date, most works rely on simple hand-crafted similarity functions and fail to generalize outside their corresponding training dataset. To address this limitation, we propose (FIRE), a framework that extracts local representations of planning problems and learns a similarity function over them. To generate the training data we introduce a novel self-supervised method that identifies similar and dissimilar pairs of local primitives from past solution paths. With these pairs, a Siamese network is trained with the contrastive loss and the similarity function is realized in the network's latent space. We evaluate FIRE on an 8-DOF manipulator in five categories of motion planning problems with sensed environments. Our experiments show that FIRE retrieves relevant experiences which can informatively guide sampling-based planners even in problems outside its training distribution, outperforming other baselines.

We present a data-driven approach for effective bin picking from clutter. Recent bin picking solutions usually lead to a direct pinch grasp on a target object without addressing any other potential contact interaction in clutter. However, appropriate physical interaction can be essential to successful singulation and subsequent secure picking, the goal of bin picking. In this work, we contribute a framework that learns physically interactive actions for object picking end-to-end from a visual input in a self-supervised manner. The learned actions enable the robot to purposefully interact with a target object by performing a digging operation through the clutter. By leveraging a fully convolutional network (FCN), we predict picking success probabilities for a set of interactive action primitives that will in turn specify an optimal action to perform. The FCN is trained in a simulated environment through trial and error. Moreover, new datasets are collected using the latest network through iterative self-supervision. Extensive real-world bin picking experiments show the effectiveness and generalizability of the approach.

Sampling-based Model Predictive Control (MPC) is a flexible control framework that can reason about non-smooth dynamics and cost functions. Recently, significant work has focused on the use of machine learning to improve the performance of MPC, often through learning or fine-tuning the dynamics or cost function. In contrast, we focus on learning to optimize more effectively. In other words, to improve the update rule within MPC. We show that this can be particularly useful in sampling-based MPC, where we often wish to minimize the number of samples for computational reasons. Unfortunately, the cost of computational efficiency is a reduction in performance; fewer samples results in noisier updates. We show that we can contend with this noise by learning how to update the control distribution more effectively and make better use of the few samples that we have. Our learned controllers are trained via imitation learning to mimic an expert which has access to substantially more samples. We test the efficacy of our approach on multiple simulated robotics tasks in sample-constrained regimes and demonstrate that our approach can outperform a MPC controller with the same number of samples.

In this work, we investigate a form of dynamic contact-rich locomotion in which a robot pushes off from obstacles in order to move through its environment. We present a reflex-based approach that switches between optimized hand-crafted reflex controllers and produces smooth and predictable motions. In contrast to previous work, our approach does not rely on periodic movements, complex models of robot and contact dynamics, or extensive hand tuning. We demonstrate the effectiveness of our approach and evaluate its performance compared to a standard model-free RL algorithm. We identify continuous clusters of similar behaviours, which allows us to successfully transfer different push-off motions directly from simulation to a physical robot without further retraining.

Navigation through uncontrolled intersections is one of the key challenges for autonomous vehicles. Identifying the subtle differences in hidden traits of other drivers can bring significant benefits when navigating in such environments. We propose an unsupervised method for inferring driver traits such as driving styles from observed vehicle trajectories. We use a variational autoencoder with recurrent neural networks to learn a latent representation of traits without any ground truth trait labels. Then, we use this trait representation to learn a policy for an autonomous vehicle to navigate through a T-intersection with deep reinforcement learning. Our pipeline enables the autonomous vehicle to adjust its actions when dealing with drivers of different traits to ensure safety and efficiency. Our method demonstrates promising performance and outperforms state-of-the-art baselines in the T-intersection scenario.

Robotic Surgical Assistants (RSAs) are commonly used to perform minimally invasive surgeries by expert surgeons. However, long procedures filled with tedious and repetitive tasks such as suturing can lead to surgeon fatigue, motivating the automation of suturing. As visual tracking of a thin reflective needle is extremely challenging, prior work has modified the needle with nonreflective contrasting paint. As a step towards automation of a suturing subtask without modifying the needle, we propose HOUSTON: Handover of Unmodified, Surgical, Tool-Obstructed Needles, a problem and algorithm that uses a learned active sensing policy with a stereo camera to iteratively localize and align the needle into a visible and accessible pose for the other gripper. To compensate for robot positioning and needle perception errors, the algorithm then executes a high-precision grasping motion that uses multiple cameras. Physical experiments with the da Vinci Research Kit (dVRK) suggest a success rate of 96.7% on needles used in training, and 75 - 92.9% on needles unseen in training. On sequential handovers, HOUSTON successfully executes 32.4 handovers on average before failure. To our knowledge, this work is the first to study handover of unmodified surgical needles. See https: / /tinyurl. com/houston-surgery for additional materials including details about offline datasets and model architectures.

Sound is an essential navigation cue that intelligent robots can leverage for localizing sound-emitting targets. This work introduces a framework for the audio-aware navigation task of mobile robots equipped with a microphone array in a complex indoor environment. The robot initialized at a random starting position has to accurately localize a distant sound source and plan an optimal path towards the sound-emitting target. Auto-encoders are used to extract implicit acoustic features that are robust against environmental noise and reverberation. The proposed framework is based on two key ideas - a sound inference module (SIM) that maps the perceived acoustic information to a given geometric map of the physical space, and a path planner that generates a path from the robot's current position to the estimated position of the sound source. Experimental results show that the SIM achieved a minimum and maximum localization error of 0.31 m and 0.70 m at a robot-source distance of 1 m and 6 m, respectively at different environmental configurations. Additionally, the proposed framework achieved a minimum and maximum reliability of 4.38 m^{-1}m^{-1} and 2.31 m^{-1}m^{-1} at a robot-source distance of 1 m and 6 m, respectively under the influence of background noise.

Manipulating an articulated object requires perceiving its kinematic hierarchy: its parts, how each can move, and how those motions are coupled. Previous work has explored perception for kinematics, but none infers a complete kinematic hierarchy on never-before-seen object instances, without relying on a schema or template. We present a novel perception system that achieves this goal. Our system infers the moving parts of an object and the kinematic couplings that relate them. To infer parts, it uses a point cloud instance segmentation neural network and to infer kinematic hierarchies, it uses a graph neural network to predict the existence, direction, and type of edges (i.e. joints) that relate the inferred parts. We train these networks using simulated scans of synthetic 3D models. We evaluate our system on simulated scans of 3D objects, and we demonstrate a proof-of-concept use of our system to drive real-world robotic manipulation.

In the peg insertion task, human pays attention to the seam between the peg and the hole and tries to fill it continuously with visual feedback. By imitating the human's behavior, we design architectures with position and orientation estimators based on the seam representation for pose alignment, which proves to be general to the unseen peg geometries. By putting the estimators into the closed-loop control with reinforcement learning, we further achieve higher or comparable success rate, efficiency, and robustness compared with the baseline methods. The policy is trained totally in simulation without any manual intervention. To achieve sim-to-real, a learnable segmentation module with automatic data collecting and labeling can be easily trained to decouple the perception and the policy, which helps the model trained in simulation quickly adapting to the real world with negligible effort. Results are presented in simulation and on a physical robot. Code, videos, and supplemental material are available at https://github.com/xieliang555/SFN.git

The ability to perceive object slip via tactile feedback enables humans to accomplish complex manipulation tasks including maintaining a stable grasp. Despite the utility of tactile information for many applications, tactile sensors have yet to be widely deployed in industrial robotics settings; part of the challenge lies in identifying slip and other events from the tactile data stream. In this paper, we present a learning-based method to detect slip using barometric tactile sensors. These sensors have many desirable properties including high durability and reliability, and are built from inexpensive, off-the-shelf components. We train a temporal convolution neural network to detect slip, achieving high detection accuracies while displaying robustness to the speed and direction of the slip motion. Further, we test our detector on two manipulation tasks involving a variety of common objects and demonstrate successful generalization to real-world scenarios not seen during training. We argue that barometric tactile sensing technology, combined with data-driven learning, is suitable for many manipulation tasks such as slip compensation.

Existing learning from demonstration algorithms usually assume access to expert demonstrations. However, this assumption is limiting in many real-world applications since the collected demonstrations may be suboptimal or even consist of failure cases. We therefore study the problem of learning from imperfect demonstrations by learning a confidence predictor. Specifically, we rely on demonstrations along with their confidence values from a different correspondent environment (source environment) to learn a confidence predictor for the environment we aim to learn a policy in (target environment-where we only have unlabeled demonstrations). We learn a common latent space through adversarial distribution matching of multi-length partial trajectories to enable the transfer of confidence across source and target environments. The learned confidence reweights the demonstrations to enable learning more from informative demonstrations and discarding the irrelevant ones. Our experiments in three simulated environments and a real robot reaching task demonstrate that our approach learns a policy with the highest expected return. We show the videos of our experiments on our website.

If robots could reliably manipulate the shape of 3D deformable objects, they could find applications in fields ranging from home care to warehouse fulfillment to surgical assistance. Analytic models of elastic, 3D deformable objects require numerous parameters to describe the potentially infinite degrees of freedom present in determining the object's shape. Previous attempts at performing 3D shape control rely on hand-crafted features to represent the object shape and require training of object-specific control models. We overcome these issues through the use of our novel DeformerNet neural network architecture, which operates on a partial-view point cloud of the object being manipulated and a point cloud of the goal shape to learn a low-dimensional representation of the object shape. This shape embedding enables the robot to learn to define a visual servo controller that provides Cartesian pose changes to the robot end-effector causing the object to deform towards its target shape. Crucially, we demonstrate both in simulation and on a physical robot that DeformerNet reliably generalizes to object shapes and material stiffness not seen during training and outperforms comparison methods for both the generic shape control and the surgical task of retraction.

This work presents the dual benefit of integrating imitation learning techniques, based on the dynamical systems formalism, with the visual servoing paradigm. On the one hand, dynamical systems allow to program additional skills without explicitly coding them in the visual servoing law, but leveraging few demonstrations of the full desired behavior. On the other, visual servoing allows to consider exteroception into the dynam-ical system architecture and be able to adapt to unexpected environment changes. The beneficial combination of the two concepts is proven by applying three existing dynamical systems methods to the visual servoing case. Simulations validate and compare the methods; experiments with a robot manipulator show the validity of the approach in a real-world scenario.

We present a novel method for generating, predicting, and using Spatiotemporal Occupancy Grid Maps (SOGM), which embed future information of dynamic scenes. Our au-tomated generation process creates groundtruth SOGMs from previous navigation data. We build on prior work to annotate lidar points based on their dynamic properties, which are then projected on time-stamped 2D grids: SOGMs. We design a 3D-2D feedforward architecture, trained to predict the future time steps of SOGMs, given 3D lidar frames as input. Our pipeline is entirely self-supervised, thus enabling lifelong learning for robots. The network is composed of a 3D back-end that extracts rich features and enables the semantic segmentation of the lidar frames, and a 2D front-end that predicts the future information embedded in the SOGMs within planning. We also design a navigation pipeline that uses these predicted SOGMs. We provide both quantitative and qualitative insights into the predictions and validate our choices of network design with a comparison to the state of the art and ablation studies.

This work aims to learn how to perform complex robot manipulation tasks that are composed of several, consecutively executed low-level sub-tasks, given as input a few visual demonstrations of the tasks performed by a person. The sub-tasks consist of moving the robot's end-effector until it reaches a sub-goal region in the task space, performing an action, and triggering the next sub-task when a pre-condition is met. Most prior work in this domain has been concerned with learning only low-level tasks, such as hitting a ball or reaching an object and grasping it. This paper describes a new neural network-based framework for learning simultaneously low-level policies as well as high-level policies, such as deciding which object to pick next or where to place it relative to other objects in the scene. A key feature of the proposed approach is that the policies are learned directly from raw videos of task demonstrations, without any manual annotation or post-processing of the data. Empirical results on object manipulation tasks with a robotic arm show that the proposed network can efficiently learn from real visual demonstrations to perform the tasks, and outperforms popular imitation learning algorithms.

This paper presents a novel graph reinforcement learning (RL) architecture to solve multi-robot task allocation (MRTA) problems that involve tasks with deadlines and workload, and robot constraints such as work capacity. While drawing motivation from recent graph learning methods that learn to solve combinatorial optimization (CO) problems such as multi-Traveling Salesman and Vehicle Routing Problems using RL, this paper seeks to provide better performance (compared to non-learning methods) and important scalability (compared to existing learning architectures) for the stated class of MRTA problems. The proposed neural architecture, called Capsule Attention-based Mechanism or CapAM acts as the policy network, and includes three main components: 1) an encoder: a Capsule Network based node embedding model to represent each task as a learnable feature vector; 2) a decoder: an attention-based model to facilitate a sequential output; and 3) context: that encodes the states of the mission and the robots. To train the CapAM model, the policy-gradient method based on REINFORCE is used. When evaluated over unseen scenar-ios, CapAM demonstrates better task completion performance and > 10 times faster decision-making compared to standard non-learning based online MRTA methods. CapAM's advantage in generalizability, and scalability to test problems of size larger than those used in training, are also successfully demonstrated in comparison to a popular approach for learning to solve CO problems, namely the purely attention mechanism.

We show that a purely tactile dextrous in-hand manipulation task with continuous regrasping, requiring permanent force closure, can be learned from scratch and executed robustly on a torque-controlled humanoid robotic hand. The task is rotating a cube without dropping it, but in contrast to OpenAI's seminal cube manipulation task [1], the palm faces downwards and no cameras but only the hand's position and torque sensing are used. Although the task seems simple, it combines for the first time all the challenges in execution as well as learning that are important for using in-hand manipulation in real-world applications. We efficiently train in a precisely modeled and identified rigid body simulation with off-policy deep reinforcement learning, significantly sped up by a domain adapted curriculum, leading to a moderate 600 CPU hours of training time. The resulting policy is robustly transferred to the real humanoid DLR Hand-II, e.g., reaching more than 46 full 2\pi2\pi rotations of the cube in a single run and allowing for disturbances like different cube sizes, hand orientation, or pulling a finger.