CoRL 2021 论文列表

Skill chaining is a promising approach for synthesizing complex behaviors by sequentially combining previously learned skills. Yet, a naive composition of skills fails when a policy encounters a starting state never seen during its training. For successful skill chaining, prior approaches attempt to widen the policy's starting state distribution. However, these approaches require larger state distributions to be covered as more policies are sequenced, and thus are limited to short skill sequences. In this paper, we propose to chain multiple policies without excessively large initial state distributions by regularizing the terminal state distributions in an adversarial learning framework. We evaluate our approach on two complex long-horizon manipulation tasks of furniture assembly. Our results have shown that our method establishes the first model-free reinforcement learning algorithm to solve these tasks; whereas prior skill chaining approaches fail. The code and videos are available at https://clvrai.com/skill-chaining

Exploiting past 3D LiDAR scans to predict future point clouds is a promising method for autonomous mobile systems to realize foresighted state estimation, collision avoidance, and planning. In this paper, we address the problem of predicting future 3D LiDAR point clouds given a sequence of past LiDAR scans. Estimating the future scene on the sensor level does not require any preceding steps as in localization or tracking systems and can be trained self-supervised. We propose an end-to-end approach that exploits a 2D range image representation of each 3D LiDAR scan and concatenates a sequence of range images to obtain a 3D tensor. Based on such tensors, we develop an encoder-decoder architecture using 3D convolutions to jointly aggregate spatial and temporal information of the scene and to predict the future 3D point clouds. We evaluate our method on multiple datasets and the experimental results suggest that our method outperforms existing point cloud prediction architectures and generalizes well to new, unseen environments without additional fine-tuning. Our method operates online and is faster than the common LiDAR frame rate of 10 Hz.

Imitating human demonstrations is a promising approach to endow robots with various manipulation capabilities. While recent advances have been made in imitation learning and batch (offline) reinforcement learning, a lack of open-source human datasets and reproducible learning methods make assessing the state of the field difficult. In this paper, we conduct an extensive study of six offline learning algorithms for robot manipulation on five simulated and three real-world multi-stage manipulation tasks of varying complexity, and with datasets of varying quality. Our study analyzes the most critical challenges when learning from offline human data for manipulation. Based on the study, we derive a series of lessons including the sensitivity to different algorithmic design choices, the dependence on the quality of the demonstrations, and the variability based on the stopping criteria due to the different objectives in training and evaluation. We also highlight opportunities for learning from human datasets, such as the ability to learn proficient policies on challenging, multi-stage tasks beyond the scope of current reinforcement learning methods, and the ability to easily scale to natural, real-world manipulation scenarios where only raw sensory signals are available. We have open-sourced our datasets and all algorithm implementations to facilitate future research and fair comparisons in learning from human demonstration data at https://arise-initiative.github.io/robomimic-web/

Efficiency in robot learning is highly dependent on hyperparameters. Robot morphology and task structure differ widely and finding the optimal setting typically requires sequential or parallel repetition of experiments, strongly increasing the interaction count. We propose a training method that only relies on a single trial by enabling agents to select and combine controller designs conditioned on the task. Our Hyperparameter Mixture Policies (HMPs) feature diverse sub-policies that vary in distribution types and parameterization, reducing the impact of design choices and unlocking synergies between low-level components. We demonstrate strong performance on continuous control tasks, including a simulated ANYmal robot, showing that HMPs yield robust, data-efficient learning.

Demonstration-guided reinforcement learning (RL) is a promising approach for learning complex behaviors by leveraging both reward feedback and a set of target task demonstrations. Prior approaches for demonstration-guided RL treat every new task as an independent learning problem and attempt to follow the provided demonstrations step-by-step, akin to a human trying to imitate a completely unseen behavior by following the demonstrator's exact muscle movements. Naturally, such learning will be slow, but often new behaviors are not completely unseen: they share subtasks with behaviors we have previously learned. In this work, we aim to exploit this shared subtask structure to increase the efficiency of demonstration-guided RL. We first learn a set of reusable skills from large offline datasets of prior experience collected across many tasks. We then propose Skill-based Learning with Demonstrations (SkiLD), an algorithm for demonstration-guided RL that efficiently leverages the provided demonstrations by following the demonstrated skills instead of the primitive actions, resulting in substantial performance improvements over prior demonstration-guided RL approaches. We validate the effectiveness of our approach on long-horizon maze navigation and complex robot manipulation tasks.

Recently, a variety of new equivariant neural network model architectures have been proposed that generalize better over rotational and reflectional symmetries than standard models. These models are relevant to robotics because many robotics problems can be expressed in a rotationally symmetric way. This paper focuses on equivariance over a visual state space and a spatial action space -- the setting where the robot action space includes a subset of $\rm{SE}(2)$. In this situation, we know a priori that rotations and translations in the state image should result in the same rotations and translations in the spatial action dimensions of the optimal policy. Therefore, we can use equivariant model architectures to make $Q$ learning more sample efficient. This paper identifies when the optimal $Q$ function is equivariant and proposes $Q$ network architectures for this setting. We show experimentally that this approach outperforms standard methods in a set of challenging manipulation problems.

Depth estimation is a safety critical and energy sensitive method for environment sensing. However, in real applications, the depth estimation may be halted at any time, due to the random interruptions or low energy capacity of battery when using powerful sensors like 3D LiDAR. To address this problem, we propose a depth estimation method that is robust to random halts and relies on energy-saving 2D LiDAR and a monocular camera. To this end, we formulate the depth estimation as an anytime problem and propose a new metric to evaluate its robustness under random interruptions. Our final model has only 2M parameters with a marginal accuracy loss compared to state-of-the-art baselines. Indeed, our experiments on NYU Depth v2 dataset show that our model is capable of processing 224$\times$224 resolution images and 2D point clouds with any computation budget larger than 6.37ms (157 FPS) and 0.2J on an NVIDIA Jetson TX2 system. Evaluations on KITTI dataset under supervised and self-supervised training show similar results.

There exist several datasets for developing self-driving car methodologies. Manually collected datasets impose inherent limitations on the variability of test cases and it is particularly difficult to acquire challenging scenarios, e.g. ones involving collisions with pedestrians. A way to alleviate this is to consider automatic generation of safety-critical scenarios for autonomous vehicle (AV) testing. Existing approaches for scenario generation use heuristic pedestrian behavior models. We instead propose a framework that can use state-of-the-art pedestrian motion models, which is achieved by reformulating the problem as learning where to place pedestrians such that the induced scenarios are collision prone for a given AV. Our pedestrian initial location model can be used in conjunction with any goal driven pedestrian model which makes it possible to challenge an AV with a wide range of pedestrian behaviors -- this ensures that the AV can avoid collisions with any pedestrian it encounters. We show that it is possible to learn a collision seeking scenario generation model when both the pedestrian and AV are collision avoiding. The initial location model is conditioned on scene semantics and occlusions to ensure semantic and visual plausibility, which increases the realism of generated scenarios. Our model can be used to test any AV model given sufficient constraints.

We propose a method that efficiently learns distributions over articulation models directly from depth images without the need to know articulation model categories a priori. By contrast, existing methods that learn articulation models from raw observations require objects to be textured, and most only predict point estimates of the model parameters. Our core contributions include a novel representation for distributions over rigid body transformations and articulation model parameters based on Screw theory, von Mises-Fisher distributions and Stiefel manifolds. Combining these concepts allows for an efficient, mathematically sound representation that inherently satisfies several constraints that rigid body transformations and articulations must adhere to. In addition, we introduce a novel deep-learning based approach, DUST-net, that efficiently learns such distributions and, hence, performs category-independent articulation model estimation while also providing model uncertainties. We evaluate our approach on two benchmarking datasets and three real-world objects and compare its performance with two current state-of-the-art methods. Our results demonstrate that DUST-net can successfully learn distributions over articulation models and their parameters for novel objects across articulation model categories with better accuracy than state-of-the-art methods.

Predicting the future trajectories of on-road vehicles is critical for autonomous driving. In this paper, we introduce a novel prediction framework called PRIME, which stands for Prediction with Model-based Planning. Unlike recent prediction works that utilize neural networks to model scene context and produce unconstrained trajectories, PRIME is designed to generate accurate and feasibility-guaranteed future trajectory predictions. PRIME guarantees the trajectory feasibility by exploiting a model-based generator to produce future trajectories under explicit constraints and enables accurate multimodal prediction by utilizing a learning-based evaluator to select future trajectories. We conduct experiments on the large-scale Argoverse Motion Forecasting Benchmark, where PRIME outperforms the state-of-the-art methods in prediction accuracy, feasibility, and robustness under imperfect tracking.

The goal of offline reinforcement learning (RL) is to find an optimal policy given prerecorded trajectories. Many current approaches customize existing off-policy RL algorithms, especially actor-critic algorithms in which policy evaluation and improvement are iterated. However, the convergence of such approaches is not guaranteed due to the use of complex non-linear function approximation and an intertwined optimization process. By contrast, we propose a simple baseline algorithm for offline RL that only performs the policy evaluation step once so that the algorithm does not require complex stabilization schemes. Since the proposed algorithm is not likely to converge to an optimal policy, it is an appropriate baseline for actor-critic algorithms that ought to be outperformed if there is indeed value in iterative optimization in the offline setting. Surprisingly, we empirically find that the proposed algorithm exhibits competitive and sometimes even state-of-the-art performance in a subset of the D4RL offline RL benchmark. This result suggests that future work is needed to fully exploit the potential advantages of iterative optimization in order to justify the reduced stability of such methods.

In this paper, we present TANDEM a real-time monocular tracking and dense mapping framework. For pose estimation, TANDEM performs photometric bundle adjustment based on a sliding window of keyframes. To increase the robustness, we propose a novel tracking front-end that performs dense direct image alignment using depth maps rendered from a global model that is built incrementally from dense depth predictions. To predict the dense depth maps, we propose Cascade View-Aggregation MVSNet (CVA-MVSNet) that utilizes the entire active keyframe window by hierarchically constructing 3D cost volumes with adaptive view aggregation to balance the different stereo baselines between the keyframes. Finally, the predicted depth maps are fused into a consistent global map represented as a truncated signed distance function (TSDF) voxel grid. Our experimental results show that TANDEM outperforms other state-of-the-art traditional and learning-based monocular visual odometry (VO) methods in terms of camera tracking. Moreover, TANDEM shows state-of-the-art real-time 3D reconstruction performance. Webpage: https://go.vision.in.tum.de/tandem

This work proposes an optimization-based manipulation planning framework where the objectives are learned functionals of signed-distance fields that represent objects in the scene. Most manipulation planning approaches rely on analytical models and carefully chosen abstractions/state-spaces to be effective. A central question is how models can be obtained from data that are not primarily accurate in their predictions, but, more importantly, enable efficient reasoning within a planning framework, while at the same time being closely coupled to perception spaces. We show that representing objects as signed-distance fields not only enables to learn and represent a variety of models with higher accuracy compared to point-cloud and occupancy measure representations, but also that SDF-based models are suitable for optimization-based planning. To demonstrate the versatility of our approach, we learn both kinematic and dynamic models to solve tasks that involve hanging mugs on hooks and pushing objects on a table. We can unify these quite different tasks within one framework, since SDFs are the common object representation. Video: https://youtu.be/ga8Wlkss7co

Semantic scene graph provides an effective way for intelligent agents to better understand the environment and it has been extensively used in many robotic applications. Existing work mainly focuses on generating the scene graph from the sensory information collected from a pre-defined path, while the environment should be exhaustively explored with a carefully designed path in order to obtain a comprehensive semantic scene graph efficiently. In this paper, we propose a new task of Embodied Semantic Scene Graph Generation, which exploits the embodiment of the intelligent agent to autonomously generate an appropriate path to explore the environment for scene graph generation. To this end, a learning framework with the paradigms of imitation learning and reinforcement learning is proposed to help the agent generate proper actions to explore the environment and the scene graph is incrementally constructed. The proposed method is evaluated on the AI2Thor environment using both the quantitative and qualitative performance indexes. Additionally, we implement the proposed method on a streaming video captioning task and promising experimental results are achieved.

In this paper, we study how robots can autonomously learn skills that require a combination of navigation and grasping. Learning robotic skills in the real world remains challenging without large scale data collection and supervision. Our aim is to devise a robotic reinforcement learning system for learning navigation and manipulation together, in an autonomous way without human intervention, enabling continual learning under realistic assumptions. Specifically, our system, ReLMM, can learn continuously on a real-world platform without any environment instrumentation, without human intervention, and without access to privileged information, such as maps, objects positions, or a global view of the environment. Our method employs a modularized policy with components for manipulation and navigation, where uncertainty over the manipulation success drives exploration for the navigation controller, and the manipulation module provides rewards for navigation. We evaluate our method on a room cleanup task, where the robot must navigate to and pick up items of scattered on the floor. After a grasp curriculum training phase, ReLMM can learn navigation and grasping together fully automatically, in around 40 hours of real-world training.

Contrary to the vast literature in modeling, perceiving, and understanding agent-object (e.g., human-object, hand-object, robot-object) interaction in computer vision and robotics, very few past works have studied the task of object-object interaction, which also plays an important role in robotic manipulation and planning tasks. There is a rich space of object-object interaction scenarios in our daily life, such as placing an object on a messy tabletop, fitting an object inside a drawer, pushing an object using a tool, etc. In this paper, we propose a unified affordance learning framework to learn object-object interaction for various tasks. By constructing four object-object interaction task environments using physical simulation (SAPIEN) and thousands of ShapeNet models with rich geometric diversity, we are able to conduct large-scale object-object affordance learning without the need for human annotations or demonstrations. At the core of technical contribution, we propose an object-kernel point convolution network to reason about detailed interaction between two objects. Experiments on large-scale synthetic data and real-world data prove the effectiveness of the proposed approach.

Robots that arrange household objects should do so according to the user's preferences, which are inherently subjective and difficult to model. We present NeatNet: a novel Variational Autoencoder architecture using Graph Neural Network layers, which can extract a low-dimensional latent preference vector from a user by observing how they arrange scenes. Given any set of objects, this vector can then be used to generate an arrangement which is tailored to that user's spatial preferences, with word embeddings used for generalisation to new objects. We develop a tidying simulator to gather rearrangement examples from 75 users, and demonstrate empirically that our method consistently produces neat and personalised arrangements across a variety of rearrangement scenarios.

Self-supervised monocular depth prediction provides a cost-effective solution to obtain the 3D location of each pixel. However, the existing approaches usually lead to unsatisfactory accuracy, which is critical for autonomous robots. In this paper, we propose FusionDepth, a novel two-stage network to advance the self-supervised monocular dense depth learning by leveraging low-cost sparse (e.g. 4-beam) LiDAR. Unlike the existing methods that use sparse LiDAR mainly in a manner of time-consuming iterative post-processing, our model fuses monocular image features and sparse LiDAR features to predict initial depth maps. Then, an efficient feed-forward refine network is further designed to correct the errors in these initial depth maps in pseudo-3D space with real-time performance. Extensive experiments show that our proposed model significantly outperforms all the state-of-the-art self-supervised methods, as well as the sparse-LiDAR-based methods on both self-supervised monocular depth prediction and completion tasks. With the accurate dense depth prediction, our model outperforms the state-of-the-art sparse-LiDAR-based method (Pseudo-LiDAR++) by more than 68% for the downstream task monocular 3D object detection on the KITTI Leaderboard.

Interacting with bins and containers is a fundamental task in robotics, making state estimation of the objects inside the bin critical. While robots often use cameras for state estimation, the visual modality is not always ideal due to occlusions and poor illumination. We introduce The Boombox, a container that uses sound to estimate the state of the contents inside a box. Based on the observation that the collision between objects and its containers will cause an acoustic vibration, we present a convolutional network for learning to reconstruct visual scenes. Although we use low-cost and low-power contact microphones to detect the vibrations, our results show that learning from multimodal data enables state estimation from affordable audio sensors. Due to the many ways that robots use containers, we believe the box will have a number of applications in robotics.