Building general-purpose robots to perform an enormous amount of tasks in a large variety of environments at the human level is notoriously complicated. It requires the robot learning to be sample-efficient, generalizable, compositional, and incremental. In this work, we introduce a systematic learning framework called SAGCI-system towards achieving these above four requirements. Our system first takes the raw point clouds gathered by the camera mounted on the robot's wrist as the inputs and produces initial modeling of the surrounding environment represented as a URDF. Our system adopts a learning-augmented differentiable simulation that loads the URDF. The robot then utilizes the interactive perception to interact with the environments to online verify and modify the URDF. Leveraging the simulation, we propose a new model-based RL algorithm combining object-centric and robot-centric approaches to efficiently produce policies to accomplish manipulation tasks. We apply our system to perform articulated object manipulation, both in the simulation and the real world. Extensive experiments demonstrate the effectiveness of our proposed learning framework. Supplemental materials and videos are available on https://sites.google.com/view/egci.

In this paper, we explore whether a robot can learn to regrasp a diverse set of objects to achieve various desired grasp poses. Regrasping is needed whenever a robot's current grasp pose fails to perform desired manipulation tasks. Endowing robots with such an ability has applications in many domains such as manufacturing or domestic services. Yet, it is a challenging task due to the large diversity of geometry in everyday objects and the high dimensionality of the state and action space. In this paper, we propose a system for robots to take partial point clouds of an object and the supporting environment as inputs and output a sequence of pick-and-place operations to transform an initial object grasp pose to the desired object grasp poses. The key technique includes a neural stable placement predictor and a regrasp graph based solution through leveraging and changing the surrounding environment. We introduce a new and challenging synthetic dataset for learning and evaluating the proposed approach. In this dataset, we show that our system is able to achieve 73.3% success rate of regrasping diverse objects.

In this paper, we explore whether a robot can learn to hang arbitrary objects onto a diverse set of supporting items such as racks or hooks. Endowing robots with such an ability has applications in many domains such as domestic services, logistics, or manufacturing. Yet, it is a challenging manipulation task due to the large diversity of geometry and topology of everyday objects. In this paper, we propose a system that takes partial point clouds of an object and a supporting item as input and learns to decide where and how to hang the object stably. Our system learns to estimate the contact point correspondences between the object and supporting item to get an estimated stable pose. We then run a deep reinforcement learning algorithm to refine the predicted stable pose. Then, the robot needs to find a collision-free path to move the object from its initial pose to stable hanging pose. To this end, we train a neural network based collision estimator that takes as input partial point clouds of the object and supporting item. We generate a new and challenging, large-scale, synthetic dataset annotated with stable poses of objects hung on various supporting items and their contact point correspondences. In this dataset, we show that our system is able to achieve a 68.3% success rate of predicting stable object poses and has a 52.1% F1 score in terms of finding feasible paths. Supplemental material and videos are available on our project webpage.

Deep reinforcement learning (DRL) algorithms have successfully been demonstrated on a range of challenging decision making and control tasks. One dominant component of recent deep reinforcement learning algorithms is the target network which mitigates the divergence when learning the Q function. However, target networks can slow down the learning process due to delayed function updates. Another dominant component especially in continuous domains is the policy gradient method which models and optimizes the policy directly. However, when Q functions are approximated with neural networks, their landscapes can be complex and therefore mislead the local gradient. In this work, we propose a self-regularized and self-guided actor-critic method. We introduce a self-regularization term within the TD-error minimization and remove the need for the target network. In addition, we propose a self-guided policy improvement method by combining policy-gradient with zero-order optimization such as the Cross Entropy Method. It helps to search for actions associated with higher Q-values in a broad neighborhood and is robust to local noise in the Q function approximation. These actions help to guide the updates of our actor network. We evaluate our method on the suite of OpenAI gym tasks, achieving or outperforming state of the art in every environment tested.

To achieve a successful grasp, gripper attributes including geometry and kinematics play a role equally important to the target object geometry. The majority of previous work has focused on developing grasp methods that generalize over novel object geometry but are specific to a certain robot hand. We propose UniGrasp, an efficient data-driven grasp synthesis method that considers both the object geometry and gripper attributes as inputs. UniGrasp is based on a novel deep neural network architecture that selects sets of contact points from the input point cloud of the object. The proposed model is trained on a large dataset to produce contact points that are in force closure and reachable by the robot hand. By using contact points as output, we can transfer between a diverse set of N-fingered robotic hands. Our model produces over 90 percent valid contact points in Top10 predictions in simulation and more than 90 percent successful grasps in the real world experiments for various known two-fingered and three-fingered grippers. Our model also achieves 93 percent and 83 percent successful grasps in the real world experiments for a novel two-fingered and five-fingered anthropomorphic robotic hand, respectively.

Endowing robots with human-like physical reasoning abilities remains challenging. We argue that existing methods often disregard spatio-temporal relations and by using Graph Neural Networks (GNNs) that incorporate a relational inductive bias, we can shift the learning process towards exploiting relations. In this work, we learn action-conditional forward dynamics models of a simulated manipulation task from visual observations involving cluttered and irregularly shaped objects. We investigate two GNN approaches and empirically assess their capability to generalize to scenarios with novel and an increasing number of objects. The first, Graph Networks (GN) based approach, considers explicitly defined edge attributes and not only does it consistently underperform an auto-encoder baseline that we modified to predict future states, our results indicate how different edge attributes can significantly influence the predictions. Consequently, we develop the Auto-Predictor that does not rely on explicitly defined edge attributes. It outperforms the baseline and the GN-based models. Overall, our results show the sensitivity of GNN-based approaches to the task representation, the efficacy of relational inductive biases and advocate choosing lightweight approaches that implicitly reason about relations over ones that leave these decisions to human designers.

We propose a method for instance-level segmentation that uses RGB-D data as input and provides detailed information about the location, geometry and number of individual objects in the scene. This level of understanding is fundamental for autonomous robots. It enables safe and robust decision-making under the large uncertainty of the real-world. In our model, we propose to use the first and second order moments of the object occupancy function to represent an object instance. We train an hourglass Deep Neural Network (DNN) where each pixel in the output votes for the 3D position of the corresponding object center and for the object's size and pose. The final instance segmentation is achieved through clustering in the space of moments. The object-centric training loss is defined on the output of the clustering. Our method outperforms the state-of-the-art instance segmentation method on our synthesized dataset. We show that our method generalizes well on real-world data achieving visually better segmentation results.