RSS 2023 论文列表

Computing optimal, collision-free trajectories for high-dimensional systems is a challenging problem. Sampling-based planners struggle with the dimensionality, whereas trajectory optimizers may get stuck in local minima due to inherent nonconvexities in the optimization landscape. The use of mixed-integer programming to encapsulate these nonconvexities and find globally optimal trajectories has recently shown great promise, thanks in part to tight convex relaxations and efficient approximation strategies that greatly reduce runtimes. These approaches were previously limited to Euclidean configuration spaces, precluding their use with mobile bases or continuous revolute joints. In this paper, we handle such scenarios by modeling configuration spaces as Riemannian manifolds, and we describe a reduction procedure for the zero-curvature case to a mixed-integer convex optimization problem. We demonstrate our results on various robot platforms, including producing efficient collision-free trajectories for a PR2 bimanual mobile manipulator.

We present NeuSE, a novel Neural SE(3)-Equivariant Embedding for objects, and illustrate how it supports object SLAM for consistent spatial understanding with longterm scene changes. NeuSE is a set of latent object embeddings created from partial object observations. It serves as a compact point cloud surrogate for complete object models, encoding full shape information while transforming SE(3)-equivariantly in tandem with the object in the physical world. With NeuSE, relative frame transforms can be directly derived from inferred latent codes. Our proposed SLAM paradigm, using NeuSE for object shape and pose characterization, can operate independently or in conjunction with typical SLAM systems. It directly infers SE(3) camera pose constraints that are compatible with general SLAM pose graph optimization, while also maintaining a lightweight object-centric map that adapts to real-world changes. Our approach is evaluated on synthetic and real-world sequences featuring changed objects and shows improved localization accuracy and change-aware mapping capability, when working either standalone or jointly with a common SLAM pipeline.

In this paper, we propose a method for estimating in-hand object poses using proprioception and tactile feedback from a bimanual robotic system. Our method addresses the problem of reducing pose uncertainty through a sequence of frictional contact interactions between the grasped objects. As part of our method, we propose 1) a tool segmentation routine that facilitates contact location and object pose estimation, 2) a loss that allows reasoning over solution consistency between interactions, and 3) a loss to promote converging to object poses and contact locations that explain the external force-torque experienced by each arm. We demonstrate the efficacy of our method in a task-based demonstration both in simulation and on a real-world bimanual platform and show significant improvement in object pose estimation over single interactions.

Motion planning seeks a collision-free path in a configuration space (C-space), representing all possible robot configurations in the environment. As it is challenging to construct a C-space explicitly for a high-dimensional robot, we generally build a graph structure called a roadmap, a discrete approximation of a complex continuous C-space, to reason about connectivity. Checking collision-free connectivity in the roadmap requires expensive edge-evaluation computations, and thus, reducing the number of evaluations has become a significant research objective. However, in practice, we often face infeasible problems: those in which there is no collision-free path in the roadmap between the start and the goal locations. Existing studies often overlook the possibility of infeasibility, becoming highly inefficient by performing many edge evaluations. In this work, we address this oversight in scenarios where a prior roadmap is available; that is, the edges of the roadmap contain the probability of being a collision-free edge learned from past experience. To this end, we propose an algorithm called iterative path and cut finding (IPC) that iteratively searches for a path and a cut in a prior roadmap to detect infeasibility while reducing expensive edge evaluations as much as possible. We further improve the efficiency of IPC by introducing a second algorithm, iterative decomposition and path and cut finding (IDPC), that leverages the fact that cut-finding algorithms partition the roadmap into smaller subgraphs. We analyze the theoretical properties of IPC and IDPC, such as completeness and computational complexity, and evaluate their performance in terms of completion time and the number of edge evaluations in large-scale simulations.

We propose DeepExplorer, a simple and lightweight metric-free exploration method for topological mapping of unknown environments. It performs task and motion planning (TAMP) entirely in image feature space. The task planner is a recurrent network using the latest image observation sequence to hallucinate a feature as the next-best exploration goal. The motion planner then utilizes the current and the hallucinated features to generate an action taking the agent towards that goal. Our novel feature hallucination enables imitation learning with deep supervision to jointly train the two planners more efficiently than baseline methods. During exploration, we iteratively call the two planners to predict the next action, and the topological map is built by constantly appending the latest image observation and action to the map and using visual place recognition (VPR) for loop closing. The resulting topological map efficiently represents an environment's connectivity and traversability, so it can be used for tasks such as visual navigation. We show DeepExplorer's exploration efficiency and strong sim2sim generalization capability on large-scale simulation datasets like Gibson and MP3D. Its effectiveness is further validated via the image-goal navigation performance on the resulting topological map. We further show its strong zero-shot sim2real generalization capability in real-world experiments. The source code is available at https://ai4ce.github.io/DeepExplorer/.

Our goal is to develop an efficient contact detection algorithm for large-scale GPU-based simulation of non-convex objects. Current GPU-based simulators such as IsaacGym [16] and Brax [11] must trade-off speed with fidelity, generality, or both when simulating non-convex objects. Their main issue lies in contact detection (CD): existing CD algorithms, such as Gilbert–Johnson–Keerthi (GJK), must trade off their computational speed with accuracy which becomes expensive as the number of collisions among non-convex objects increases. We propose a data-driven approach for CD, whose accuracy depends only on the quality and quantity of offline dataset rather than online computation time. Unlike GJK, our method inherently has a uniform computational flow, which facilitates efficient GPU usage based on advanced compilers such as XLA (Accelerated Linear Algebra) [2]. Further, we offer a data-efficient solution by learning the patterns of colliding local crop object shapes, rather than global object shapes which are harder to learn. We demonstrate our approach improves the efficiency of existing CD methods by a factor of 5-10 for nonconvex objects with comparable accuracy. Using the previous work on contact resolution for a neural-network-based contact detector [23], we integrate our CD algorithm into the open-source GPU-based simulator, Brax, and show that we can improve the efficiency over IsaacGym and generality over standard Brax. We highly recommend the videos of our simulator included in the supplementary materials. (https://sites.google.com/view/locc-rss2023/home)

Robotic grasping is an essential and fundamental task and has been studied extensively over the past several decades. Traditional work analyzes physical models of the objects and computes force-closure grasps. Such methods require pre-knowledge of the complete 3D model of an object, which can be hard to obtain. Recently with significant progress in machine learning, data-driven methods have dominated the area. Although impressive improvements have been achieved, those methods require a vast amount of training data and suffer from limited generalizability. In this paper, we propose a novel two-stage approach to predicting and synthesizing grasping poses directly from the point cloud of an object without database knowledge or learning. Firstly, multiple superquadrics are recovered at different positions within the object, representing the local geometric features of the object surface. Subsequently, our algorithm exploits the tri-symmetry feature of superquadrics and synthesizes a list of antipodal grasps from each recovered superquadric. An evaluation model is designed to assess and quantify the quality of each grasp candidate. The grasp candidate with the highest score is then selected as the final grasping pose. We conduct experiments on isolated and packed scenes to corroborate the effectiveness of our method. The results indicate that our method demonstrates competitive performance compared with the state-of-the-art without the need for either a full model or prior training.

Legged robots have enormous potential in their range of capabilities, from navigating unstructured terrains to high-speed running. However, these capabilities bring with them difficult control problems, and designing controllers for highly agile dynamic motions remains a substantial challenge for roboticists. Reinforcement learning (RL) offers a promising data-driven approach for automatically training such controllers. However, exploration in these high-dimensional, underactuated systems remains a significant hurdle for enabling legged robots to learn performant, naturalistic, and versatile agility skills. We propose a framework for training complex robotic skills by transferring experience from existing controllers to jumpstart learning new tasks. To leverage controllers we can acquire in practice, we design this framework to be flexible in terms of their source---that is, the controllers may have been optimized for a different objective under different dynamics, or may require different knowledge of the surroundings---and thus may be highly suboptimal for the target task. We show that our method enables learning complex agile jumping behaviors, navigating to goal locations while walking on hind legs, and adapting to new environments. We also demonstrate that the agile behaviors learned in this way are graceful and safe enough to deploy in the real world.

Fine manipulation tasks, such as threading cable ties or slotting a battery, are notoriously difficult for robots because they require precision, careful coordination of contact forces, and closed-loop visual feedback. Performing these tasks typically requires high-end robots, accurate sensors, or careful calibration, which can be expensive and difficult to set up. Can learning enable low-cost and imprecise hardware to perform these fine manipulation tasks? We present a low-cost system that performs end-to-end imitation learning directly from real demonstrations, collected with a custom teleoperation interface. Imitation learning, however, presents its own challenges, particularly in high-precision domains: the error of the policy can compound over time, drifting out of the training distribution. To address this challenge, we develop a simple yet novel algorithm Action Chunking with Transformers (ACT) which reduces the effective horizon by predicting actions in chunks. This allows us to learn difficult tasks such as opening a translucent condiment cup and slotting a battery with 80-90% success, with only 10 minutes worth of demonstration data. Project website: https://tonyzhaozh.github.io/aloha/

Recent work in visual representation learning for robotics demonstrates the viability of learning from large video datasets of humans performing everyday tasks. Leveraging methods such as masked autoencoding and contrastive learning, these representations exhibit strong transfer to policy learning for visuomotor control. But, robot learning encompasses a diverse set of problems beyond control including grasp affordance prediction, language-conditioned imitation learning, and intent scoring for human-robot collaboration, amongst others. First, we demonstrate that existing representations yield inconsistent results across these tasks: masked autoencoding approaches pick up on low-level spatial features at the cost of high-level semantics, while contrastive learning approaches capture the opposite. We then introduce Voltron, a framework for language-driven representation learning from human videos and associated captions. Voltron trades off language-conditioned visual reconstruction to learn low-level visual patterns, and visually-grounded language generation to encode high-level semantics. We also construct a new evaluation suite spanning five distinct robot learning problems – a unified platform for holistically evaluating visual representations for robotics. Through comprehensive, controlled experiments across all five problems, we find that Voltron’s language-driven representations outperform the prior state-of-the-art, especially on targeted problems requiring higher-level features.

Dexterous manipulation has been a long-standing challenge in robotics. While machine learning techniques have shown some promise, results have largely been currently limited to simulation. This can be mostly attributed to the lack of suitable hardware. In this paper, we present LEAP Hand, a low-cost dexterous and anthropomorphic hand for machine learning research. In contrast to previous hands, LEAP Hand has a novel kinematic structure that allows maximal dexterity regardless of finger pose. LEAP Hand is low-cost and can be assembled in 4 hours at a cost of 2000 USD from readily available parts. It is capable of consistently exerting large torques over long durations of time. We show that LEAP Hand can be used to perform several manipulation tasks in the real world---from visual teleoperation to learning from passive video data and sim2real. LEAP Hand significantly outperforms its closest competitor Allegro Hand in all our experiments while being 1/8th of the cost. We release the URDF model, 3D CAD files, tuned simulation environment, and a development platform with useful APIs on our website at https://leap-hand.github.io/ .

The field of Learning from Demonstration enables end-users, who are not robotics experts, to shape robot behavior. However, using human demonstrations to teach robots to solve long-horizon problems by leveraging the hierarchical structure of the task is still an unsolved problem. Prior work has yet to show that human users can provide sufficient demonstrations in novel domains without showing the demonstrators explicit teaching strategies for each domain. In this work, we investigate whether non-expert demonstrators can generalize robot teaching strategies to provide necessary and sufficient demonstrations to robots zero-shot in novel domains. We find that increasing participant experience with providing demonstrations improves their demonstration's degree of sub-task abstraction (p<.001), teaching efficiency (p<.001), and sub-task redundancy (p<.05) in novel domains, allowing generalization in robot teaching. Our findings demonstrate for the first time that non-expert demonstrators can transfer knowledge from a series of training experiences to novel domains without the need for explicit instruction, such that they can provide necessary and sufficient demonstrations when programming robots to complete task and motion planning problems.

Reasoning over the interplay between object deformation and force transmission through contact is central to the manipulation of compliant objects. In this paper, we propose Neural Deforming Contact Field (NDCF), a representation that jointly models object deformations and contact patches from visuo-tactile feedback using implicit representations. Representing the object geometry and contact with the environment implicitly allows a single model to predict contact patches of varying complexity. Additionally, learning geometry and contact simultaneously allows us to enforce physical priors, such as ensuring contacts lie on the surface of the object. We propose a neural network architecture to learn a NDCF, and train it using simulated data. We then demonstrate that the learned NDCF transfers directly to the real-world without the need for fine-tuning. We benchmark our proposed approach against a baseline representing geometry and contact patches with point clouds. We find that NDCF performs better on simulated data and in transfer to the real-world. More details and video results can be found at https://www.mmintlab.com/ndcf/.

Localization for autonomous robots in prior maps is crucial for their functionality. This paper offers a solution to this problem for indoor environments called InstaLoc, which operates on an individual lidar scan to localize it within a prior map. We draw on inspiration from how humans navigate and position themselves by recognizing the layout of distinctive objects and structures. Mimicking the human approach, InstaLoc identifies and matches object instances in the scene with those from a prior map. As far as we know, this is the first method to use panoptic segmentation directly inferring on 3D lidar scans for indoor localization. InstaLoc operates through two networks based on spatially sparse tensors to directly infer dense 3D lidar point clouds. The first network is a panoptic segmentation network that produces object instances and their semantic classes. The second smaller network produces a descriptor for each object instance. A consensus based matching algorithm then matches the instances to the prior map and estimates a six degrees of freedom (DoF) pose for the input cloud in the prior map. InstaLoc utilizes two efficient networks, requires only one to two hours of training on a mobile GPU, and runs in real-time at 1 Hz. Our method achieves between two and four times more detections when localizing, as compared to baseline methods, and achieves higher precision on these detections.

Robotic assembly is a longstanding challenge, requiring contact-rich interaction and high precision and accuracy. Many applications also require adaptivity to diverse parts, poses, and environments, as well as low cycle times. In other areas of robotics, simulation is a powerful tool to develop algorithms, generate datasets, and train agents. However, simulation has had a more limited impact on assembly. We present IndustReal, a set of algorithms, systems, and tools that solve assembly tasks in simulation with reinforcement learning (RL) and successfully achieve policy transfer to the real world. Specifically, we propose 1) simulation-aware policy updates, 2) signed-distance-field rewards, and 3) sampling-based curricula for robotic RL agents. We use these algorithms to enable robots to solve contact-rich pick, place, and insertion tasks in simulation. We then propose 4) a policy-level action integrator to minimize error at policy deployment time. We build and demonstrate a real-world robotic assembly system that uses the trained policies and action integrator to achieve repeatable performance in the real world. Finally, we present hardware and software tools that allow other researchers to reproduce our system and results. For videos and additional details, please see our project website at https://sites.google.com/nvidia.com/industreal.

This paper presents a novel approach for optical flow control of Micro Air Vehicles (MAVs). The task is challenging due to the nonlinearity of optical flow observables. Our proposed Incremental Nonlinear Dynamic Inversion (INDI) control scheme incorporates an efficient data-driven method to address the nonlinearity. It directly estimates the inverse of the time-varying control effectiveness in real-time, eliminating the need for the constant assumption and avoiding high computation in traditional INDI. This approach effectively handles fast-changing system dynamics commonly encountered in optical flow control, particularly height-dependent changes. We demonstrate the robustness and efficiency of the proposed control scheme in numerical simulations and also real-world flight tests: multiple landings of an MAV on a static and flat surface with various tracking setpoints, hovering and landings on moving and undulating surfaces. Despite being challenged with the presence of noisy optical flow estimates and the lateral and vertical movement of the landing surfaces, the MAV is able to successfully track or land on the surface with an exponential decay of both height and vertical velocity at almost the same time, as desired.

Object goal navigation is an important problem in Embodied AI that involves guiding the agent to navigate to an instance of the object category in an unknown environment---typically an indoor scene. Unfortunately, current state-of-the-art methods for this problem rely heavily on data-driven approaches, e.g., end-to-end reinforcement learning, imitation learning, and others. Moreover, such methods are typically costly to train and difficult to debug, leading to a lack of transferability and explainability. Inspired by recent successes in combining classical and learning methods, we present a modular and training-free solution, which embraces more classic approaches, to tackle the object goal navigation problem. Our method builds a structured scene representation based on the classic visual simultaneous localization and mapping (V-SLAM) framework. We then inject semantics into geometric-based frontier exploration to reason about promising areas to search for a goal object. Our structured scene representation comprises a 2D occupancy map, semantic point cloud, and spatial scene graph. Our method propagates semantics on the scene graphs based on language priors and scene statistics to introduce semantic knowledge to the geometric frontiers. With injected semantic priors, the agent can reason about the most promising frontier to explore. The proposed pipeline shows strong experimental performance for object goal navigation on the Gibson benchmark dataset, outperforming the previous state-of-the-art. We also perform comprehensive ablation studies to identify the current bottleneck in the object navigation task.

Reinforcement learning has shown great potential in solving complex tasks when large amounts of data can be generated with little effort. In robotics, one approach to generate training data builds on simulations or models. However, for many tasks, such as with complex soft robots, devising such models is substantially more challenging. Recent successes in soft robotics indicate that employing complex robots can lead to performance boosts. Here, we leverage the imbalance in complexity of the dynamics to learn more sample-efficiently. We (i) abstract the task into distinct components, (ii) off-load the simple dynamics parts into the simulation, and (iii) multiply these virtual parts to generate more data in hindsight. Our new method, Hindsight States (HiS), uses this data and selects the most useful transitions for training. It can be used with an arbitrary off-policy algorithm. We validate our method on several challenging simulated tasks and demonstrate that it improves learning both on its own and when combined with an existing hindsight algorithm, Hindsight Experience Replay (HER). Finally, we evaluate HiS on a physical system and show that it boosts performance on a complex table tennis task with a muscular robot.

Visual-inertial odometry (VIO) is the most common approach for estimating the state of autonomous micro aerial vehicles using only onboard sensors. Existing methods improve VIO performance by including a dynamics model in the estimation pipeline. However, such methods degrade in the presence of low-fidelity vehicle models and continuous external disturbances, such as wind. Our proposed method, HDVIO, overcomes these limitations by using a hybrid dynamics model that combines a point-mass vehicle model with a learning-based component that captures complex aerodynamic effects. HDVIO estimates the external force and the full robot state by leveraging the discrepancy between the actual motion and the predicted motion of the hybrid dynamics model. Our hybrid dynamics model uses a history of thrust and IMU measurements to predict the vehicle dynamics. To demonstrate the performance of our method, we present results on both public and novel drone dynamics datasets and show real-world experiments of a quadrotor flying in strong winds up to 25 km/h. The results show that our approach improves the motion and external force estimation compared to the state-of-the-art by up to 33% and 40%, respectively. Furthermore, differently from existing methods, we show that it is possible to predict the vehicle dynamics accurately while having no explicit knowledge of its full state.