ICRA 2021 Papers

Learning for model based control can be sample-efficient and generalize well, however successfully learning models and controllers that represent the problem at hand can be challenging for complex tasks. Using inaccurate models for learning can lead to sub-optimal solutions that are unlikely to perform well in practice. In this work, we present a learning approach which iterates between model learning and data collection and leverages forward model prediction error for learning control. We show how using the controller’s prediction as input to a forward model can create a differentiable connection between the controller and the model, allowing us to formulate a loss in the state space. This lets us include forward model prediction error during controller learning and we show that this creates a loss objective that significantly improves learning on different motor control tasks. We provide empirical and theoretical results that show the benefits of our method and present evaluations in simulation for learning control on a 7 DoF manipulator and an underactuated 12 DoF quadruped. We show that our approach successfully learns controllers for challenging motor control tasks involving contact switching.

Many studies have focused on Virtual Reality (VR) frameworks for remotely controlling robotic systems. Although VR systems have been used to teleoperate robots in simple scenarios, their effectiveness in terms of accuracy, speed, and usability has not been rigorously evaluated for complex tasks that require accurate trajectories. In this work, an Enhanced Virtual Reality (EVR) framework for robotic teleoperation is evaluated to assess if it can be efficiently used in complex tasks that require accurate control of the robotic end-effector. The environment and the employed robot are captured using RGB-D cameras, while the remote user controls the motion of the robot with VR controllers. The captured data are transmitted and reconstructed in 3D so as to allow the remote user to monitor the task execution progress in real time, using a VR headset. The EVR system is compared with two other interface alternatives: i) teleoperation in pure VR (the model of the robot is rendered with respect to its real joint states), and ii) teleoperation in EVRR (the model of the robot is superimposed on the real robot). The results show that pure point cloud interfaces suffer from visualization issues, reducing the effectiveness of the robot teleoperation. However, the accuracy and user experience can be greatly improved by including the robot model.

This paper proposes a state estimator for legged robots operating in slippery environments. An Invariant Extended Kalman Filter (InEKF) is implemented to fuse inertial and velocity measurements from a tracking camera and leg kinematic constraints. The misalignment between the camera and the robot-frame is also modeled thus enabling auto-calibration of camera pose. The leg kinematics based velocity measurement is formulated as a right-invariant observation. Nonlinear observability analysis shows that other than the rotation around the gravity vector and the absolute position, all states are observable except for some singular cases. Discrete observability analysis demonstrates that our filter is consistent with the underlying nonlinear system. An online noise parameter tuning method is developed to adapt to the highly time-varying camera measurement noise. The proposed method is experimentally validated on a Cassie bipedal robot walking over slippery terrain. A video for the experiment can be found at https://youtu.be/VIqJL0cUr7s.

We introduce a methodology to compute the inverse kinematics for concentric tube continuum robots from a desired shape as input. We demonstrate that it is possible to accurately learn joint parameters using neural networks for a discrete point-wise shape representation with different discretization. In comparison to a vanilla numerical method, the learning-based method is preferred in terms of accuracy in joint space and computation. Representing the shape with up to 20 equidistant points, a shape-to-joint inverse kinematics with errors of 2.22° and 1.45 mm is obtained. Further, we extend the shape-to-joint inverse kinematics to image-to-joint inverse kinematics utilizing multi-view images as shape representation. This image-based method achieves errors of 6.02° and 2.76 mm. Both approaches, i.e., shape-to-joint and image-to-joint, result in higher accuracy compared to the learning-based state-of-the-art approach which only considers the tip pose.

The combinatorics inherent to the issue of planning legged locomotion can be addressed by decomposing the problem: first, select a guide path abstracting the contacts with a heuristic model; then compute the contact sequence to balance the robot gait along the guide path. While several models have been proposed to compute such a path, none have yet managed to efficiently capture the complexity of legged locomotion on arbitrary terrain. In this paper, we present a novel method to automatically build a local controller, or steering method, to generate a guide path along which a feasible contact sequence can be built. Our reinforcement learning approach is coupled with a geometric condition for feasibility during the training, which improves the convergence rate without incurring a loss in generality. We have designed a dedicated environment along with an associated reward function to run a classical reinforcement learning algorithm that computes the steering method. The policy takes a target velocity and a local heightmap of the terrain around the robot as inputs, and steers the path where new contacts should be created. It is then coupled with a contact generator that creates the contacts to support the robot movement. We show that the trained policy has an improved generalization and higher success rate at generating feasible contact plans than previous approaches. As a result, this policy can be used with a path planning algorithm to navigate complex environments.

Recently, autonomous driving has made substantial progress in addressing the most common traffic scenarios like intersection navigation and lane changing. However, most of these successes have been limited to scenarios with well-defined traffic rules and require minimal negotiation with other vehicles. In this paper, we introduce a previously unconsidered, yet everyday, high-conflict driving scenario requiring negotiations between agents of equal rights and priorities. There exists no centralized control structure and we do not allow communications. Therefore, it is unknown if other drivers are willing to cooperate, and if so to what extent. We train policies to robustly negotiate with opposing vehicles of an unobservable degree of cooperativeness using multi-agent reinforcement learning (MARL). We propose Discrete Asymmetric Soft Actor-Critic (DASAC), a maximum- entropy off-policy MARL algorithm allowing for centralized training with decentralized execution. We show that using DASAC we are able to successfully negotiate and traverse the scenario considered over 99% of the time. Our agents are robust to an unknown timing of opponent decisions, an unobservable degree of cooperativeness of the opposing vehicle, and previously unencountered policies. Furthermore, they learn to exhibit human-like behaviors such as defensive driving, anticipating solution options and interpreting the behavior of other agents.

Rearranging and manipulating deformable objects such as cables, fabrics, and bags is a long-standing challenge in robotic manipulation. The complex dynamics and high-dimensional configuration spaces of deformables, compared to rigid objects, make manipulation difficult not only for multi-step planning, but even for goal specification. Goals cannot be as easily specified as rigid object poses, and may involve complex relative spatial relations such as "place the item inside the bag". In this work, we develop a suite of simulated benchmarks with 1D, 2D, and 3D deformable structures, including tasks that involve image-based goal-conditioning and multi-step deformable manipulation. We propose embedding goal-conditioning into Transporter Networks, a recently proposed model architecture for learning robotic manipulation that rearranges deep features to infer displacements that can represent pick and place actions. We demonstrate that goal-conditioned Transporter Networks enable agents to manipulate deformable structures into flexibly specified configurations without test-time visual anchors for target locations. We also significantly extend prior results using Transporter Networks for manipulating deformable objects by testing on tasks with 2D and 3D deformables. Supplementary material is available at https://berkeleyautomation.github.io/bags/.

Modeling dynamics of deformable linear objects (DLOs), such as cables, hoses, sutures, and catheters, is an important and challenging problem for many robotic manipulation applications. In this paper, we propose the first method to model and learn full 3D dynamics of DLOs from data. Our approach is capable of capturing the complex twisting and bending dynamics of DLOs and allows local effects to propagate globally. To this end, we adapt the interaction network (IN) dynamics learning method for capturing the interaction between neighboring segments in a DLO and augment it with a recurrent model for propagating interaction effects along the length of a DLO. For learning twisting and bending dynamics in 3D, we also introduce a new suitable representation of DLO segments and their relationships. Unlike the original IN method, our model learns to propagate the effects of local interaction between neighboring segments to each segment in the chain within a single time step, without the need for iterated propagation steps. Evaluation of our model with synthetic and newly collected real-world data shows better accuracy and generalization in short-term and long- term predictions than the current state of the art. We further integrate our learned model in a model predictive control scheme and use it to successfully control the shape of a DLO. Our implementation is available at https://gitsvn-nt.oru.se/ammlab–public/in–bilstm.

Many robotics applications require interest points that are highly repeatable under varying viewpoints and lighting conditions. However, this requirement is very challenging as the environment changes continuously and indefinitely, leading to appearance changes of interest points with respect to time. This paper proposes to predict the repeatability of an interest point as a function of time, which can tell us the lifespan of the interest point considering daily or seasonal variation. The repeatability predictor (RP) is formulated as a regressor trained on repeated interest points from multiple viewpoints over a long period of time. Through comprehensive experiments, we demonstrate that our RP can estimate when a new interest point is repeated, and also highlight an insightful analysis about this problem. For further comparison, we apply our RP to the map summarization under visual localization framework, which builds a compact representation of the full context map given the query time. The experimental result shows a careful selection of potentially repeatable interest points predicted by our RP can significantly mitigate the degeneration of localization accuracy from map summarization.

Most existing approaches for visual localization either need a detailed 3D model of the environment or, in the case of learning-based methods, must be retrained for each new scene. This can either be very expensive or simply impossible for large, unknown environments, for example in search-and-rescue scenarios. Although there are learning-based approaches that operate scene-agnostically, the generalization capability of these methods is still outperformed by classical approaches. In this paper, we present an approach that can generalize to new scenes by applying specific changes to the model architecture, including an extended regression part, the use of hierarchical correlation layers, and the exploitation of scale and uncertainty information. Our approach outperforms the 5-point algorithm using SIFT features on equally big images and additionally surpasses all previous learning-based approaches that were trained on different data. It is also superior to most of the approaches that were specifically trained on the respective scenes. We also evaluate our approach in a scenario with only very few reference images, showing that under such more realistic conditions our learning-based approach considerably exceeds both existing learning-based and classical methods.

In this paper, we introduce a reinforcement learning approach utilizing a novel topology-based information gain metric for directing the next best view of a noisy 3D sensor. The metric combines the disjoint sections of an observed surface to focus on high-detail features such as holes and concave sections. Experimental results show that our approach can aid in establishing the placement of a robotic sensor to optimize the information provided by its streaming point cloud data. Furthermore, a labeled dataset of 3D objects, a CAD design for a custom robotic manipulator, and software for the transformation, union, and registration of point clouds has been publicly released to the research community.

We propose a multi-time-scale predictive representation learning method to efficiently learn robust driving policies in an offline manner that generalize well to novel road geometries, and damaged and distracting lane conditions which are not covered in the offline training data. We show that our proposed representation learning method can be applied easily in an offline (batch) reinforcement learning setting demonstrating the ability to generalize well and efficiently under novel conditions compared to standard batch RL methods. Our proposed method utilizes training data collected entirely offline in the real-world which removes the need of intensive online explorations that impede applying deep reinforcement learning on real-world robot training. Various experiments were conducted in both simulator and real-world scenarios for the purpose of evaluation and analysis of our proposed claims.

We present a reinforcement learning-based solution to autonomously race on a miniature race car platform. We show that a policy that is trained purely in simulation using a relatively simple vehicle model, including model randomization, can be successfully transferred to the real robotic setup. We achieve this by using a novel policy output regularization approach and a lifted action space which enables smooth actions but still aggressive race car driving. We show that this regularized policy does outperform the Soft Actor Critic (SAC) baseline method, both in simulation and on the real car, but it is still outperformed by a Model Predictive Controller (MPC) state-of-the-art method. The refinement of the policy with three hours of real-world interaction data allows the reinforcement learning policy to achieve lap times similar to the MPC controller while reducing track constraint violations by 50%.

State-of-the-art reinforcement learning (RL) algorithms suffer from high sample complexity, particularly in the sparse reward case. A popular strategy for mitigating this problem is to learn control policies by imitating a set of expert demonstrations. The drawback of such approaches is that an expert needs to produce demonstrations, which may be costly in practice. To address this shortcoming, we propose Probabilistic Planning for Demonstration Discovery (P2D2), a technique for automatically discovering demonstrations without access to an expert. We formulate discovering demonstrations as a search problem and leverage widely-used planning algorithms such as Rapidly-exploring Random Tree to find demonstration trajectories. These demonstrations are used to initialize a policy, then refined by a generic RL algorithm. We provide theoretical guarantees of P2D2 finding successful trajectories, as well as bounds for its sampling complexity. We experimentally demonstrate the method outperforms classic and intrinsic exploration RL techniques in a range of classic control and robotics tasks, requiring only a fraction of exploration samples and achieving better asymptotic performance.

In shared autonomy, the user input is blended with the assistive motion to accomplish a task where the user goal is typically unknown to the robot. Transparency between the human and robot is essential for effective collaboration. Prior works have provided methods for the robot to infer the user goal; however, they are usually dependent on the distance between the robot and object, which may not be directly associated with the real-time user control intention and thus cause low control feelings. Here, we propose a real-time goal prediction method driven by assistive motion generated by learning from demonstration (LfD) allowing more reactive assistive behaviors. This LfD-generated assistive motion is blended with the user input based on goal predictions to achieve targeted tasks. The LfD policy was learned offline and used with different users. To evaluate our proposed method, we compared it with a state-of-the-art Partially Observable Markov Decision Process (POMDP) based method using a distance cost, and a direct control method (i.e., joystick). A pilot study (N = 6) was conducted to control a 6-DoF Kinova Mico robotic arm to carry out three tasks: (1) reaching-and-grasping, (2) pouring, and (3) object-returning with the three control methods. We used both objective and subjective measures in the comparative study. Results show that our method has the shortest task completion time, the lowest amount of joystick control inputs among all three control methods, as well as a significantly lower angular difference between the user input and assistive motion compared to the POMDP-based method. Besides, it obtains the highest subjective score in the user preference and perceived speed ratings, and the second-highest in the control feeling and the robot did what I wanted ratings.

We propose a novel planning technique for satisfying tasks specified in temporal logic in partially revealed environments. We define high-level actions derived from the environment and the given task itself, and estimate how each action contributes to progress towards completing the task. As the map is revealed, we estimate the cost and probability of success of each action from images and an encoding of that action using a trained neural network. These estimates guide search for the minimum-expected-cost plan within our model. Our learned model is structured to generalize across environments and task specifications without requiring retraining. We demonstrate an improvement in total cost in both simulated and real-world experiments compared to a heuristic-driven baseline.

Estimating a depth map from a single RGB image has been investigated widely for localization, mapping, and 3- dimensional object detection. Recent studies on a single-view depth estimation are mostly based on deep Convolutional neural Networks (ConvNets) which require a large amount of training data paired with densely annotated labels. Depth annotation tasks are both expensive and inefficient, so it is inevitable to leverage RGB images which can be collected very easily to boost the performance of ConvNets without depth labels. However, most self-supervised learning algorithms are focused on capturing the semantic information of images to improve the performance in classification or object detection, not in depth estimation. In this paper, we show that existing self- supervised methods do not perform well on depth estimation and propose a gradient-based self-supervised learning algorithm with momentum contrastive loss to help ConvNets extract the geometric information with unlabeled images. As a result, the network can estimate the depth map accurately with a relatively small amount of annotated data. To show that our method is independent of the model structure, we evaluate our method with two different monocular depth estimation algorithms. Our method outperforms the previous state-of-the- art self-supervised learning algorithms and shows the efficiency of labeled data in triple compared to random initialization on the NYU Depth v2 dataset.

Whole-body optimizers have been successful at automatically computing complex dynamic locomotion behaviors. However they are often limited to offline planning as they are computationally too expensive to replan with a high frequency. Simpler models are then typically used for online replanning. In this paper we present a method to generate whole body movements in real-time for locomotion tasks. Our approach consists in learning a centroidal neural network that predicts the desired centroidal motion given the current state of the robot and a desired contact plan. The network is trained using an existing whole body motion optimizer. Our approach enables to learn with few training samples dynamic motions that can be used in a complete whole-body control framework at high frequency, which is usually not attainable with typical full-body optimizers. We demonstrate our method to generate a rich set of walking and jumping motions on a real quadruped robot.

Moving in dynamic pedestrian environments is one of the important requirements for autonomous mobile robots. We present a model-based reinforcement learning approach for robots to navigate through crowded environments. The navigation policy is trained with both real interaction data from multi-agent simulation and virtual data from a deep transition model that predicts the evolution of surrounding dynamics of mobile robots. A reward function considering social conventions is designed to guide the training of the policy. Specifically, the policy model takes laser scan sequence and robot’s own state as input and outputs steering command. The laser sequence is further transformed into stacked local obstacle maps disentangled from robot’s ego motion to separate the static and dynamic obstacles, simplifying the model training. We observe that the policy using our method can be trained with significantly less real interaction data in simulator but achieve similar level of success rate in social navigation tasks compared with other methods. Experiments are conducted in multiple social scenarios both in simulation and on real robots, the learned policy can guide the robots to the final targets successfully in a socially compliant manner. Code is available at https://github.com/YuxiangCui/model-based-social-navigation.