Though robot learning is often formulated in terms of discrete-time Markov decision processes (MDPs), physical robots require near-continuous multiscale feedback control. Machines operate on multiple asynchronous sensing modalities, each with different frequencies, e.g., video frames at 30Hz, proprioceptive state at 100Hz, force-torque data at 500Hz, etc. While the classic approach is to batch observations into fixed-time windows then pass them through feed-forward encoders (e.g., with deep networks), we show that there exists a more elegant approach - one that treats policy learning as modeling latent state dynamics in continuous-time. Specifically, we present InFuser, a unified architecture that trains continuous time-policies with Neural Controlled Differential Equations (CDEs). InFuser evolves a single latent state representation over time by (In)tegrating and (Fus)ing multi-sensory observations (arriving at different frequencies), and inferring actions in continuous-time. This enables policies that can react to multi-frequency multi-sensory feedback for truly end-to-end visuomotor control, without discrete-time assumptions. Behavior cloning experiments demonstrate that InFuser learns robust policies for dynamic tasks (e.g., swinging a ball into a cup) notably outperforming several baselines in settings where observations from one sensing modality can arrive at much sparser intervals than others.

While both navigation and manipulation are chal-lenging topics in isolation, many tasks require the ability to both navigate and manipulate in concert. To this end, we propose a mobile manipulation system that leverages novel navigation and shape completion methods to manipulate an object with a mobile robot. Our system utilizes uncertainty in the initial estimation of a manipulation target to calculate a predicted next-best-view. Without the need of localization, the robot then uses the predicted panoramic view at the next-best-view location to navigate to the desired location, capture a second view of the object, create a new model that predicts the shape of object more accurately than a single image alone, and uses this model for grasp planning. We show that the system is highly effective for mobile manipulation tasks through simulation experiments using real world data, as well as ablations on each component of our system.

We consider the problem of learning useful robotic skills from previously collected offline data without access to manually specified rewards or additional online exploration, a setting that is becoming increasingly important for scaling robot learning by reusing past robotic data. In particular, we propose the objective of learning a functional understanding of the environment by learning to reach any goal state in a given dataset. We employ goal-conditioned Q-learning with hindsight relabeling and develop several techniques that enable training in a particularly challenging offline setting. We find that our method can operate on high-dimensional camera images and learn a variety of skills on real robots that generalize to previously unseen scenes and objects. We also show that our method can learn to reach long-horizon goals across multiple episodes through goal chaining, and learn rich representations that can help with downstream tasks through pre-training or auxiliary objectives.

We propose a vision-based architecture search algorithm for robot manipulation learning, which discovers interactions between low dimension action inputs and high dimensional visual inputs. Our approach automatically designs architectures while training on the task – discovering novel ways of combining and attending image feature representations with actions as well as features from previous layers. The obtained new architectures demonstrate better task success rates, in some cases with a large margin, compared to a recent high performing baseline. Our real robot experiments also confirm that it improves grasping performance by 6%. This is the first approach to demonstrate a successful neural architecture search and attention connectivity search for a real-robot task.

Deep Q learning (DQN) has enabled robot agents to accomplish vision based tasks that seemed out of reach. Despite recent success stories, there are still several sources of computational complexity that challenge the performance of DQN. We place the focus on vision manipulation tasks, where the correct action selection is often predicated on a small number of pixels. We observe that in some of these tasks DQN does not converge to the optimal Q function, and their values do not separate well optimal and suboptimal actions. In consequence, the policies obtained with DQN tend to be brittle and manifest a low success rate, especially in long horizon tasks. In this work we show the benefits of Reward Machines (RMs) for Deep Q learning (DQRM) in vision based robot manipulation tasks. Reward machines decompose the task at an abstract level, inform the agent about their current stage along task completion, and guide them via dense rewards. We show that RMs help DQN learn the optimal Q values in each abstract state. Their policies are more robust, manifest higher success rate, and are learned with fewer training steps compared with DQN. The benefits of RMs are more evident in long-horizon tasks, where we show that DQRM is able to learn good-quality policies with six times times fewer training steps than DQN, even when this is equipped with dense reward shaping.

In this work, we present an effective multi-view approach to closed-loop end-to-end learning of precise manipulation tasks that are 3D in nature. Our method learns to accomplish these tasks using multiple statically placed but uncalibrated RGB camera views without building an explicit 3D representation such as a pointcloud or voxel grid. This multi-camera approach achieves superior task performance on difficult stacking and insertion tasks compared to single-view baselines. Single view robotic agents struggle from occlusion and challenges in estimating relative poses between points of interest. While full 3D scene representations (voxels or pointclouds) are obtainable from registered output of multiple depth sensors, several challenges complicate operating off such explicit 3D representations. These challenges include imperfect camera calibration, poor depth maps due to object properties such as reflective surfaces, and slower inference speeds over 3D representations compared to 2D images. Our use of static but uncalibrated cameras does not require camera-robot or camera-camera calibration making the proposed approach easy to setup and our use of sensor dropout during training makes it resilient to the loss of camera-views after deployment.

We present a new paradigm for Neural ODE algorithms, called ODEtoODE, where time-dependent parameters of the main flow evolve according to a matrix flow on the orthogonal group O(d). This nested system of two flows, where the parameter-flow is constrained to lie on the compact manifold, provides stability and effectiveness of training and solves the gradient vanishing-explosion problem which is intrinsically related to training deep neural network architectures such as Neural ODEs. Consequently, it leads to better downstream models, as we show on the example of training reinforcement learning policies with evolution strategies, and in the supervised learning setting, by comparing with previous SOTA baselines. We provide strong convergence results for our proposed mechanism that are independent of the width of the network, supporting our empirical studies. Our results show an intriguing connection between the theory of deep neural networks and the field of matrix flows on compact manifolds.

This work provides an architecture that incorporates depth and tactile information to create rich and accurate 3D models useful for robotic manipulation tasks. This is accomplished through the use of a 3D convolutional neural network (CNN). Offline, the network is provided with both depth and tactile information and trained to predict the object's geometry, thus filling in regions of occlusion. At runtime, the network is provided a partial view of an object. Tactile information is acquired to augment the captured depth information. The network can then reason about the object's geometry by utilizing both the collected tactile and depth information. We demonstrate that even small amounts of additional tactile information can be incredibly helpful in reasoning about object geometry. This is particularly true when information from depth alone fails to produce an accurate geometric prediction. Our method is benchmarked against and outperforms other visual-tactile approaches to general geometric reasoning. We also provide experimental results comparing grasping success with our method.

This work provides a framework for a workspace aware online grasp planner. This framework greatly improves the performance of standard online grasp planning algorithms by incorporating a notion of reachability into the online grasp planning process. Offline, a database of hundreds of thousands of unique end-effector poses were queried for feasibility. At runtime, our grasp planner uses this database to bias the hand towards reachable end-effector configurations. The bias keeps the grasp planner in accessible regions of the planning scene so that the resulting grasps are tailored to the situation at hand. This results in a higher percentage of reachable grasps, a higher percentage of successful grasp executions, and a reduced planning time. We also present experimental results using simulated and real environments.

This work provides an architecture to enable robotic grasp planning via shape completion. Shape completion is accomplished through the use of a 3D convolutional neural network (CNN). The network is trained on our own new open source dataset of over 440,000 3D exemplars captured from varying viewpoints. At runtime, a 2.5D pointcloud captured from a single point of view is fed into the CNN, which fills in the occluded regions of the scene, allowing grasps to be planned and executed on the completed object. Runtime shape completion is very rapid because most of the computational costs of shape completion are borne during offline training. We explore how the quality of completions vary based on several factors. These include whether or not the object being completed existed in the training data and how many object models were used to train the network. We also look at the ability of the network to generalize to novel objects allowing the system to complete previously unseen objects at runtime. Finally, experimentation is done both in simulation and on actual robotic hardware to explore the relationship between completion quality and the utility of the completed mesh model for grasping.

This paper presents a deep learning architecture for detecting the palm and fingertip positions of stable grasps directly from partial object views. The architecture is trained using RGBD image patches of fingertip and palm positions from grasps computed on complete object models using a grasping simulator. At runtime, the architecture is able to estimate grasp quality metrics without the need to explicitly calculate the given metric. This ability is useful as the exact calculation of these quality functions is impossible from an incomplete view of a novel object without any tactile feedback. This architecture for grasp quality prediction provides a framework for generalizing grasp experience from known to novel objects.