RSS 2022 论文列表

Humans are capable of completing a range of challenging manipulation tasks that require reasoning jointly over modalities such as vision, touch, and sound. Moreover, many such tasks are partially-observed; for example, taking a notebook out of a backpack will lead to visual occlusion and require reasoning over the history of audio or tactile information. While robust tactile sensing can be costly to capture on robots, microphones near or on a robot's gripper are a cheap and easy way to acquire audio feedback of contact events, which can be a surprisingly valuable data source for perception in the absence of vision. Motivated by the potential for sound to mitigate visual occlusion, we aim to learn a set of challenging partially-observed manipulation tasks from visual and audio inputs. Our proposed system learns these tasks by combining offline imitation learning from a modest number of tele-operated demonstrations and online finetuning using human provided interventions. In a set of simulated tasks, we find that our system benefits from using audio, and that by using online interventions we are able to improve the success rate of offline imitation learning by ~20%. Finally, we find that our system can complete a set of challenging, partially-observed tasks on a Franka Emika Panda robot, like extracting keys from a bag, with a 70% success rate, 50% higher than a policy that does not use audio.

Differential Dynamic Programming (DDP) is an efficient trajectory optimization algorithm relying on second-order approximations of a system's dynamics and cost function, and has recently been applied to optimize systems with time-invariant parameters. Prior works include system parameter estimation and identifying the optimal switching time between modes of hybrid dynamical systems. This paper generalizes previous work by proposing a general parameterized optimal control objective and deriving a parametric version of DDP, titled Parameterized Differential Dynamic Programming (PDDP). A rigorous convergence analysis of the algorithm is provided, and PDDP is shown to converge to a minimum of the cost regardless of initialization. The effects of varying the optimization to more effectively escape local minima are analyzed. Experiments are presented applying PDDP on multiple robotics systems to solve model predictive control (MPC) and moving horizon estimation (MHE) tasks simultaneously. Finally, PDDP is used to determine the optimal transition point between flight regimes of a complex urban air mobility (UAM) class vehicle exhibiting multiple phases of flight.

Quadratic programming (QP) has become a core modelling component in the modern engineering toolkit. This is particularly true for simulation, planning and control in robotics. Yet, modern numerical solvers have not reached the level of efficiency and reliability required in practical applications where speed, robustness, and accuracy are all necessary. In this work, we introduce a few variations of the well-established augmented Lagrangian method, specifically for solving QPs, which include heuristics for improving practical numerical performances. Those variants are embedded within an open-source software which includes an efficient C++ implementation, a modular API, as well as best-performing heuristics for our test-bed. Relying on this framework, we present a benchmark studying the practical performances of modern optimization solvers for convex QPs on generic and complex problems of the literature as well as on common robotic scenarios. This benchmark notably highlights that this approach outperforms modern solvers in terms of efficiency, accuracy and robustness for small to medium-sized problems, while remaining competitive for higher dimensions.

Maintaining an up-to-date map to reflect recent changes in the scene is very important, particularly in situations involving repeated traversals by a robot operating in an environment over an extended period. Undetected changes may cause a deterioration in map quality, leading to poor localization, inefficient operations, and lost robots. Volumetric methods, such as truncated signed distance functions (TSDFs), have quickly gained traction due to their real-time production of a dense and detailed map, though map updating in scenes that change over time remains a challenge. We propose a framework that introduces a novel probabilistic object state representation to track object pose changes in semi-static scenes. The representation jointly models a stationarity score and a TSDF change measure for each object. A Bayesian update rule that incorporates both geometric and semantic information is derived to achieve consistent online map maintenance. To extensively evaluate our approach alongside the state-of-the-art, we release a novel real-world dataset in a warehouse environment. We also evaluate on the public ToyCar dataset. Our method outperforms state-of-the-art methods on the reconstruction quality of semi-static environments.

In this paper, we propose an optimization based SLAM approach to simultaneously optimize the robot trajectory and the occupancy map using 2D laser scans (and odometry) information. The key novelty is that the robot poses and the occupancy map are optimized together, which is significantly different from existing occupancy mapping strategies where the robot poses need to be obtained first before the map can be estimated. In our formulation, the map is represented as a continuous occupancy map where each 2D point in the environment has a corresponding evidence value. The Occupancy-SLAM problem is formulated as an optimization problem where the variables include all the robot poses and the occupancy values at the selected discrete grid cell nodes. We propose a variation of Gauss-Newton method to solve this new formulated problem, obtaining the optimized occupancy map and robot trajectory together with their uncertainties. Our algorithm is an offline approach since it is based on batch optimization and the number of variables involved is large. Evaluations using simulations and publicly available practical 2D laser datasets demonstrate that the proposed approach can estimate the maps and robot trajectories more accurately than the state-of-the-art techniques, when a relatively accurate initial guess is provided to our algorithm. The video shows the convergence process of the proposed Occupancy-SLAM and comparison of results to Cartographer can be found at https://youtu.be/4oLyVEUC4iY.

Learning from demonstration (LfD) seeks to democratize robotics by enabling non-experts to intuitively program robots to perform novel skills through human task demonstration. Yet, LfD is challenging under a task and motion planning setting which requires hierarchical abstractions. Prior work has studied mechanisms for eliciting demonstrations that include hierarchical specifications of task and motion, via keyframes [1] or hierarchical task network specifications [2]. However, such prior works have not examined whether non-roboticist end-users are capable of providing such hierarchical demonstrations without explicit training from a roboticist showing how to teach each task [3]. To address the limitations and assumptions of prior work, we conduct two novel human-subjects experiments to answer (1) what are the necessary conditions to teach users through hierarchy and task abstractions and (2) what instructional information or feedback is required to support users to learn to program robots effectively to solve novel tasks. Our first experiment shows that fewer than half (35.71%) of our subjects provide demonstrations with sub-task abstractions when not primed. Our second experiment demonstrates that users fail to teach the robot correctly when not shown a video demonstration of an expert’s teaching strategy for the exact task that the subject is training. Not even showing the video of an analogue task was sufficient. These experiments reveal the need for fundamentally different approaches in LfD which can allow end-users to teach generalizable long-horizon tasks to robots without the need to be coached by experts at every step.

In this paper we present a control barrier function-based (CBF) resilience controller that provides resilience in a multi-robot network to adversaries. Previous approaches provide resilience by virtue of specific linear combinations of multiple control constraints. These combinations can be difficult to find and are sensitive to the addition of new constraints. Unlike previous approaches, the proposed CBF provides network resilience and is easily amenable to multiple other control constraints, such as collision and obstacle avoidance. The inclusion of such constraints is essential in order to implement a resilience controller on realistic robot platforms. We demonstrate the viability of the CBF-based resilience controller on real robotic systems through case studies on a multi-robot flocking problem in cluttered environments with the presence of adversarial robots.

This paper considers policy-centric optimal motion planning with limited reaction time. The motion planning queries are determined by their goal regions and cost functionals, and are generated over time from a distribution. Once a new query is requested, the robot needs to quickly generate a motion planner which can steer the robot to the goal region while minimizing a cost functional. We develop a meta-learning-based algorithm to compute a meta value function, which can be fast adapted using a small number of samples of a new query. Simulations on a unicycle are conducted to evaluate the developed algorithm and show the anytime property of the proposed algorithm.

Self-occlusion is challenging for cloth manipulation, as it makes it difficult to estimate the full state of the cloth. Ideally, a robot trying to unfold a crumpled or folded cloth should be able to reason about the cloth's occluded regions. For example, suppose that a robot is trying to unfold a square towel with a corner folded beneath, the robot can successfully unfold it only if it knows the existence of occluded corner. We leverage recent advances in pose estimation for cloth to build a system that uses explicit occlusion reasoning to unfold a crumpled cloth. Specifically, we first learn a model to reconstruct the mesh of the cloth. However, the model will likely have errors due to the complexities of the cloth configurations and due to ambiguities from occlusions. Our main insight is that we can further refine the predicted reconstruction by performing test-time finetuning with self-supervised losses. The obtained reconstructed mesh allows us to use a mesh-based dynamics model for planning while reasoning about occlusions. We evaluate our system both on cloth flattening as well as on cloth canonicalization, in which the objective is to manipulate the cloth into a canonical pose. Our experiments show that our method significantly outperforms prior methods that do not explicitly account for occlusions or perform test-time optimization. Videos and visualizations can be found on our website: https://sites.google.com/view/occlusion-reason/home.

Communication is a hallmark of intelligence. In this work, we present MIRROR, an approach to (i) quickly learn human models from human demonstrations, and (ii) use the models for subsequent communication planning in assistive shared-control settings. MIRROR is inspired by social projection theory, which hypothesizes that humans use self-models to understand others. Likewise, MIRROR leverages self-models learned using reinforcement learning to bootstrap human modeling. Experiments with simulated humans show that this approach leads to rapid learning and more robust models compared to existing behavioral cloning and state-of-the-art imitation learning methods. We also present a human-subject study using the CARLA simulator which shows that (i) MIRROR is able to scale to complex domains with high-dimensional observations and complicated world physics and (ii) provides effective assistive communication that enabled participants to drive more safely in adverse weather conditions.

In multi-agent settings, game theory is a natural framework for describing the strategic interactions of agents whose objectives depend upon one another's behavior. Trajectory games capture these complex effects by design. In competitive settings, this makes them a more faithful interaction model than traditional "predict then plan" approaches. However, current game-theoretic planning methods have important limitations. In this work, we propose two main contributions. First, we introduce an offline training phase which reduces the online computational burden of solving trajectory games. Second, we formulate a lifted game which allows players to optimize multiple candidate trajectories in unison and thereby construct more competitive "mixed" strategies. We validate our approach on a number of experiments using the pursuit-evasion game "tag".

Gradient-based approaches in reinforcement learning (RL) have achieved tremendous success in learning policies for autonomous vehicles. While the performance of these approaches warrants real-world adoption, these policies lack interpretability, limiting deployability in the safety-critical and legally-regulated domain of autonomous driving (AD). AD requires interpretable and verifiable control policies that maintain high performance. We propose Interpretable Continuous Control Trees (ICCTs), a tree-based model that can be optimized via modern, gradient-based, RL approaches to produce high-performing, interpretable policies. The key to our approach is a procedure for allowing direct optimization in a sparse decision-tree-like representation. We validate ICCTs against baselines across six domains, showing that ICCTs are capable of learning interpretable policy representations that parity or outperform baselines by up to 33% in AD scenarios while achieving a 300x-600x reduction in the number of policy parameters against deep learning baselines. Furthermore, we demonstrate the interpretability and utility of our ICCTs through a 14-car physical robot demonstration.

For autonomous quadruped robot navigation in various complex environments, a typical SOTA system is composed of four main modules -- mapper, global planner, local planner, and command-tracking controller -- in a hierarchical manner. In this paper, we build a robust and safe local planner which is designed to generate a velocity plan to track a coarsely planned path from the global planner. Previous works used waypoint-based methods (e.g. Proportional-Differential control and pure pursuit) which simplify the path tracking problem to local point-goal navigation. However, they suffer from frequent collisions in geometrically complex and narrow environments because of two reasons; the global planner uses a coarse and inaccurate model and the local planner is unable to track the global plan sufficiently well. Currently, deep learning methods are an appealing alternative because they can learn safety and path feasibility from experience more accurately. However, existing deep learning methods are not capable of planning for a long horizon. In this work, we propose a learning-based fully autonomous navigation framework composed of three innovative elements: a learned forward dynamics model (FDM), an online sampling-based model-predictive controller, and an informed trajectory sampler (ITS). Using our framework, a quadruped robot can autonomously navigate in various complex environments without a collision and generate a smoother command plan compared to the baseline method. Furthermore, our method can reactively handle unexpected obstacles on the planned path and avoid them.

Simultaneous localization and mapping (SLAM) in the deformable environment has encountered several barricades. One of them is the lack of a global registration technique. Thus current SLAM systems heavily rely on template based methods. We propose KernelGPA, a novel global registration technique to bridge the gap. We define nonrigid transformations using a kernel method, and show that the principal axes of the map can be solved globally in closed-form, up to a global scale ambiguity along each axis. We propose to solve both the global scale ambiguity and rigid poses in a unified optimization framework, yielding a cost that can be readily incorporated in sensor fusion frameworks. We demonstrate the registration performance of KernelGPA using various datasets, with a special focus on computerized tomography (CT) registration. We release our code and data to foster future research in this direction.

This paper tackles the task of goal-conditioned dynamic manipulation of deformable objects. This task is highly challenging due to its complex dynamics (introduced by object deformation and high-speed action) and strict task requirements (defined by a precise goal specification). To address these challenges, we present Iterative Residual Policy (IRP), a general learning framework applicable to repeatable tasks with complex dynamics. IRP learns an implicit policy via delta dynamics -- instead of modeling the entire dynamical system and inferring actions from that model, IRP learns delta dynamics that predict the effects of delta action on the previously-observed trajectory. When combined with adaptive action sampling, the system can quickly optimize its actions online to reach a specified goal. We demonstrate the effectiveness of IRP on two tasks: whipping a rope to hit a target point and swinging a cloth to reach a target pose. Despite being trained only in simulation on a fixed robot setup, IRP is able to efficiently generalize to noisy real-world dynamics, new objects with unseen physical properties, and even different robot hardware embodiments, demonstrating its excellent generalization capability relative to alternative approaches.

A robot's deployment environment often involves perceptual changes that differ from what it has experienced during training. Standard practices such as data augmentation attempt to bridge this gap by augmenting source images in an effort to extend the support of the training distribution to better cover what the agent might experience at test time. In many cases, however, it is impossible to know test-time distribution-shift a priori, making these schemes infeasible. In this paper, we introduce a general approach, called Invariance through Latent Alignment (ILA), that improves the test-time performance of a visuomotor control policy in deployment environments with unknown perceptual variations. ILA performs unsupervised adaptation at deployment-time by matching the distribution of latent features on the target domain to the agent's prior experience, without relying on paired data. Although simple, we show that this idea leads to surprising improvements on a variety of challenging adaptation scenarios, including changes in lighting conditions, the content in the scene, and camera poses. We present results on calibrated control benchmarks in simulation --the distractor control suite-- and a physical robot under a sim-to-real setup.

3D scene graphs have recently emerged as a powerful high-level representation of 3D environments. A 3D scene graph models the environment as a layered graph where nodes represent spatial concepts at multiple levels of abstraction (from low-level geometry to high-level semantics including objects, places, rooms, buildings, etc.) and edges represent relations between concepts. While 3D scene graphs can serve as an advanced “mental model” for robots, how to build such a rich representation in real-time is still uncharted territory. This paper describes a real-time Spatial Perception System, a suite of algorithms to build a 3D scene graph from sensor data in real-time. Our first contribution is to develop real-time algorithms to incrementally construct the layers of a scene graph as the robot explores the environment; these algorithms build a local ESDF around the current robot trajectory estimate, extract a topological map of places from the ESDF, and then segment the places into rooms using an approach inspired by community-detection techniques. Our second contribution is to investigate loop closure detection and optimization in 3D scene graphs. We show that 3D scene graphs allow defining hierarchical descriptors for place recognition; our descriptors capture statistics across layers in the scene graph, ranging from low-level visual appearance, to summary statistics about objects and places. We then propose the first algorithm to optimize a 3D scene graph in response to loop closures; our approach relies on embedded deformation graphs to simultaneously correct all layers of the scene graph. We implement the proposed system into a highly parallelized architecture, named Hydra, that combines fast early and mid-level perception processes with slower high-level perception. We evaluate Hydra on simulated and real data and show it is able to reconstruct 3D scene graphs with an accuracy comparable with batch offline methods, while running online.

We approach the problem of learning from watching humans in the wild. While traditional approaches in Imitation and Reinforcement Learning are promising for learning in the real world, they are either sample inefficient or are constrained to lab settings. Meanwhile, there has been a lot of success in processing passive, unstructured human data. We propose tackling this problem via an efficient one-shot robot learning algorithm, centered around learning from a third person perspective. We call our method WHIRL: In the Wild Human-Imitated Robot Learning. In WHIRL, we aim to use human videos to gather a prior over the intent of the demonstrator, advances in computer vision, and use this to initialize our agent's policy. We introduce an efficient real-world policy learning scheme, that improves over the human prior using interactions. Our key contributions are a simple sampling-based policy optimization approach, a novel objective function for aligning human and robot videos as well as an exploration method to boost sample efficiency. We show, one-shot, generalization and success in real world settings, including 20 different manipulation tasks in the wild.

A motion-based control interface promises flexible robot operations in dangerous environments by combining user intuitions with the robot's motor capabilities. However, designing a motion interface for non-humanoid robots, such as quadrupeds or hexapods, is not straightforward because different dynamics and control strategies govern their movements. We propose a novel motion control system that allows a human user to operate various motor tasks seamlessly on a quadrupedal robot. We first retarget the captured human motion into the corresponding robot motion with proper semantics using supervised learning and post-processing techniques. Then we apply the motion imitation learning with curriculum learning to develop a control policy that can track the given retargeted reference. We further improve the performance of both motion retargeting and motion imitation by training a set of experts. As we demonstrate, a user can execute various motor tasks using our system, including standing, sitting, tilting, manipulating, walking, and turning, on simulated and real quadrupeds. We also conduct a set of studies to analyze the performance gain induced by each component.