Robotic in-hand manipulation has been a long-standing challenge due to the complexity of modelling hand and object in contact and of coordinating finger motion for complex manipulation sequences. To address these challenges, the majority of prior work has either focused on model-based, low-level controllers or on model-free deep reinforcement learning that each have their own limitations. We propose a hierarchical method that relies on traditional, model-based controllers on the low-level and learned policies on the mid-level. The low-level controllers can robustly execute different manipulation primitives (reposing, sliding, flipping). The mid-level policy orchestrates these primitives. We extensively evaluate our approach in simulation with a 3-fingered hand that controls three degrees of freedom of elongated objects. We show that our approach can move objects between almost all the possible poses in the workspace while keeping them firmly grasped. We also show that our approach is robust to inaccuracies in the object models and to observation noise. Finally, we show how our approach generalizes to objects of other shapes.
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.
Understanding dynamic 3D environment is crucial for robotic agents and many other applications. We propose a novel neural network architecture called $MeteorNet$ for learning representations for dynamic 3D point cloud sequences. Different from previous work that adopts a grid-based representation and applies 3D or 4D convolutions, our network directly processes point clouds. We propose two ways to construct spatiotemporal neighborhoods for each point in the point cloud sequence. Information from these neighborhoods is aggregated to learn features per point. We benchmark our network on a variety of 3D recognition tasks including action recognition, semantic segmentation and scene flow estimation. MeteorNet shows stronger performance than previous grid-based methods while achieving state-of-the-art performance on Synthia. MeteorNet also outperforms previous baseline methods that are able to process at most two consecutive point clouds. To the best of our knowledge, this is the first work on deep learning for dynamic raw point cloud sequences.
When teams of robots collaborate to complete a task, communication is often necessary. Like humans, robot teammates should implicitly communicate through their actions: but interpreting our partner's actions is typically difficult, since a given action may have many different underlying reasons. Here we propose an alternate approach: instead of not being able to infer whether an action is due to exploration, exploitation, or communication, we define separate roles for each agent. Because each role defines a distinct reason for acting (e.g., only exploit, only communicate), teammates now correctly interpret the meaning behind their partner's actions. Our results suggest that leveraging and alternating roles leads to performance comparable to teams that explicitly exchange messages.
Rich haptic sensory feedback in response to user interactions is desirable for an effective, immersive virtual reality or teleoperation system. However, this feedback depends on material properties and user interactions in a complex, non-linear manner. Therefore, it is challenging to model the mapping from material and user interactions to haptic feedback in a way that generalizes over many variations of the user's input. Current methodologies are typically conditioned on user interactions, but require a separate model for each material. In this paper, we present a learned action-conditional model that uses data from a vision-based tactile sensor (GelSight) and user's action as input. This model predicts an induced acceleration that could be used to provide haptic vibration feedback to a user. We trained our proposed model on a publicly available dataset (Penn Haptic Texture Toolkit) that we augmented with GelSight measurements of the different materials. We show that a unified model over all materials outperforms previous methods and generalizes to new actions and new instances of the material categories in the dataset.
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.
Reinforcement Learning (RL) of contact-rich manipulation tasks has yielded impressive results in recent years. While many studies in RL focus on varying the observation space or reward model, few efforts focused on the choice of action space (e.g. joint or end-effector space, position, velocity, etc.). However, studies in robot motion control indicate that choosing an action space that conforms to the characteristics of the task can simplify exploration and improve robustness to disturbances. This paper studies the effect of different action spaces in deep RL and advocates for Variable Impedance Control in End-effector Space (VICES) as an advantageous action space for constrained and contact-rich tasks. We evaluate multiple action spaces on three prototypical manipulation tasks: Path Following (task with no contact), Door Opening (task with kinematic constraints), and Surface Wiping (task with continuous contact). We show that VICES improves sample efficiency, maintains low energy consumption, and ensures safety across all three experimental setups. Further, RL policies learned with VICES can transfer across different robot models in simulation, and from simulation to real for the same robot. Further information is available at https://stanfordvl.github.io/vices.
Contact-rich manipulation tasks in unstructured environments often require both haptic and visual feedback. It is non-trivial to manually design a robot controller that combines these modalities which have very different characteristics. While deep reinforcement learning has shown success in learning control policies for high-dimensional inputs, these algorithms are generally intractable to deploy on real robots due to sample complexity. In this work, we use self-supervision to learn a compact and multimodal representation of our sensory inputs, which can then be used to improve the sample efficiency of our policy learning. Evaluating our method on a peg insertion task, we show that it generalizes over varying geometries, configurations, and clearances, while being robust to external perturbations. We also systematically study different self-supervised learning objectives and representation learning architectures. Results are presented in simulation and on a physical robot.