RSS 2023 Papers

We present our general-purpose mobile manipulation system consisting of a custom robot platform and key algorithms spanning perception and planning. To extensively test the system in the wild and benchmark its performance, we choose a grocery shopping scenario in an actual, unmodified grocery store. We derive key performance metrics from detailed robot log data collected during six week-long field tests, spread across 18 months. These objective metrics, gained from complex yet repeatable tests, drive the direction of our research efforts and let us continuously improve our system’s performance. We find that thorough end-to-end system-level testing of a complex mobile manipulation system can serve as a reality-check for state-of-the-art methods in robotics. This effectively grounds robotics research efforts in real world needs and challenges, which we deem highly useful for the advancement of the field. To this end, we share our key insights and takeaways to inspire and accelerate similar system-level research projects.

Automating warehouse operations can reduce logistics overhead costs, ultimately driving down the final price for consumers, increasing the speed of delivery, and enhancing the resiliency to workforce fluctuations. The past few years have seen increased interest in automating such repeated tasks but mostly in controlled settings. Tasks such as picking objects from unstructured, cluttered piles have only recently become robust enough for large-scale deployment with minimal human intervention. This paper demonstrates a large-scale package manipulation from unstructured piles in Amazon Robotics' Robot Induction (Robin) fleet, which utilizes a pick success predictor trained on real production data. Specifically, the system was trained on over 394K picks. It is used for singulating up to 5~million packages per day and has manipulated over 200~million packages during this paper's evaluation period. The developed learned pick quality measure ranks various pick alternatives in real-time and prioritizes the most promising ones for execution. The pick success predictor aims to estimate from prior experience the success probability of a desired pick by the deployed industrial robotic arms in cluttered scenes containing deformable and rigid objects with partially known properties. It is a shallow machine learning model, which allows us to evaluate which features are most important for the prediction. An online pick ranker leverages the learned success predictor to prioritize the most promising picks for the robotic arm, which are then assessed for collision avoidance. This learned ranking process is demonstrated to overcome the limitations and outperform the performance of manually engineered and heuristic alternatives. To the best of the authors' knowledge, this paper presents the first large-scale deployment of learned pick quality estimation methods in a real production system.

Robots may benefit from large language models (LLMs), which have demonstrated strong reasoning capabilities across various domains. This demonstration includes several systems based on recent methods that integrate LLMs on robots: SayCan, Socratic Models, Inner Monologue, and Code as Policies. While each algorithm highlights a different mode of grounding, they all share a common system-level structure in that they use LLMs to take as input natural language instructions and generate robot plans in the form of step-by-step procedures or code. This structure provides several practical perks for demonstration in that (i) we can use existing video chat interfaces to instruct the robot by typing commands and broadcasting its movements in action via video streaming, (ii) one can seamlessly switch between interfaces that communicate with different robots, and (iii) this can all be done remotely on a laptop, where the robots on real hardware can be held on standby in the lab ready to run on command. Our tentative plan is to show at least one system running on real hardware remotely -- Inner Monologue or Code as Policies, and solicit task instructions from a live audience. Time-permitting we may also demonstrate the other systems available to run on real hardware. Otherwise, we will present recorded videos of past runs. We will link to open-source code, and conclude with a discussion of open research questions in the area.

In recent years, mobile robot navigation approaches have become increasingly important due to various application areas ranging from healthcare to warehouse logistics. In particular, Deep Reinforcement Learning approaches have gained popularity for robot navigation but are not easily accessible to non-experts and complex to develop. In recent years, efforts have been made to make these sophisticated approaches accessible to a wider audience. In this paper, we present Arena-Web, a web-based development and evaluation suite for developing, training, and testing DRL-based navigation planners for various robotic platforms and scenarios. The interface is designed to be intuitive and engaging to appeal to non-experts and make the technology accessible to a wider audience. With Arena-Web and its interface, training and developing Deep Reinforcement Learning agents is simplified and made easy without a single line of code. The web-app is free to use and openly available under the link stated in the supplementary materials.

Deep reinforcement learning is a promising approach to learning policies in unstructured environments. Due to its sample inefficiency, though, deep RL applications have primarily focused on simulated environments. In this work, we demonstrate that the recent advancements in machine learning algorithms and libraries combined with careful MDP formulation lead to learning quadruped locomotion in only 20 minutes in the real world. We evaluate our approach on several indoor and outdoor terrains that are known to be challenging for classical, model-based controllers and observe that the robot consistently learns a walking gait on all of these terrains. Finally, we evaluate our design decisions in a simulated environment.

We describe a system for deep reinforcement learning of robotic manipulation skills applied to a large-scale real-world task: sorting recyclables and trash in office buildings. Real-world deployment of deep RL policies requires not only effective training algorithms, but the ability to bootstrap real-world training and enable broad generalization. To this end, our system combines scalable deep RL from real-world data with bootstrapping from training in simulation, and incorporates auxiliary inputs from existing computer vision systems as a way to boost generalization to novel objects, while retaining the benefits of end-to-end training. We analyze the tradeoffs of different design decisions in our system, and present a large-scale empirical validation that includes training on real-world data gathered over the course of 24 months of experimentation, across a fleet of 23 robots in three office buildings, with a total training set of 9527 hours of robotic experience. Our final validation also consists of 4800 evaluation trials across 240 waste station configurations, in order to evaluate in detail the impact of the design decisions in our system, the scaling effects of including more real-world data, and the performance of the method on novel objects.

Scaling to arbitrarily large bundle adjustment problems requires data and compute to be distributed across multiple devices. Centralized methods in prior works are only able to solve small or medium size problems due to overhead in computation and communication. In this paper, we present a fully decentralized method that alleviates computation and communication bottlenecks to solve arbitrarily large bundle adjustment problems. We achieve this by reformulating the reprojection error and deriving a novel surrogate function that decouples optimization variables from different devices. This function makes it possible to use majorization minimization techniques and reduces bundle adjustment to independent optimization subproblems that can be solved in parallel. We further apply Nesterov's acceleration and adaptive restart to improve convergence while maintaining its theoretical guarantees. Despite limited peer-to-peer communication, our method has provable convergence to first-order critical points under mild conditions. On extensive benchmarks with public datasets, our method converges much faster than decentralized baselines with similar memory usage and communication load. Compared to centralized baselines using a single device, our method, while being decentralized, yields more accurate solutions with significant speedups of up to 953.7x over Ceres and 174.6x over DeepLM. Code: https://github.com/facebookresearch/DABA.

This paper reports a novel result: with proper robot models on matrix Lie groups, one can formulate the kinodynamic motion planning problem for rigid body systems as \emph{exact} polynomial optimization problems that can be relaxed as semidefinite programming (SDP). Due to the nonlinear rigid body dynamics, the motion planning problem for rigid body systems is nonconvex. Existing global optimization-based methods do not properly deal with the configuration space of the 3D rigid body; thus, they do not scale well to long-horizon planning problems. We use Lie groups as the configuration space in our formulation and apply the variational integrator to formulate the forced rigid body systems as quadratic polynomials. Then we leverage Lasserre's hierarchy to obtain the globally optimal solution via SDP. By constructing the motion planning problem in a sparse manner, the results show that the proposed algorithm has \emph{linear} complexity with respect to the planning horizon. This paper demonstrates the proposed method can provide rank-one optimal solutions at relaxation order two for most of the testing cases of 1) 3D drone landing using the full dynamics model and 2) inverse kinematics for serial manipulators.

Task allocation can enable effective coordination of multi-robot teams to accomplish tasks that are intractable for individual robots. However, existing approaches to task allocation often assume that task requirements or reward functions are known and explicitly specified by the user. In this work, we consider the challenge of forming effective coalitions for a given heterogeneous multi-robot team when task reward functions are unknown. To this end, we first formulate a new class of problems, dubbed COncurrent Constrained Online optimization of Allocation (COCOA). The COCOA problem requires online optimization of coalitions such that the unknown rewards of all the tasks are simultaneously maximized using a given multi-robot team with constrained resources. To address the COCOA problem, we introduce an online optimization algorithm, named Concurrent Multi-Task Adaptive Bandits (CMTAB), that leverages and builds upon continuum-armed bandit algorithms. Experiments involving detailed numerical simulations and a simulated emergency response task reveal that CMTAB can effectively trade-off exploration and exploitation to simultaneously and efficiently optimize the unknown task rewards while respecting the team's resource constraints.

Building 3D maps of the environment is central to robot navigation, planning, and interaction with objects in a scene. Most existing approaches that integrate semantic concepts with 3D maps largely remain confined to the closed-set setting: they can only reason about a finite set of concepts, pre-defined at training time. To address this issue, we propose ConceptFusion, a scene representation that is: (i) fundamentally open-set, enabling reasoning beyond a closed set of concepts (ii) inherently multi-modal, enabling a diverse range of possible queries to the 3D map, from language, to images, to audio, to 3D geometry, all working in concert. ConceptFusion leverages the open-set capabilities of today’s foundation models that have been pre-trained on internet-scale data to reason about concepts across modalities such as natural language, images, and audio. We demonstrate that pixel-aligned open-set features can be fused into 3D maps via traditional SLAM and multi-view fusion approaches. This enables effective zero-shot spatial reasoning, not needing any additional training or finetuning, and retains long-tailed concepts better than supervised approaches, outperforming them by more than 40% margin on 3D IoU. We extensively evaluate ConceptFusion on a number of real-world datasets, simulated home environments, a real-world tabletop manipulation task, and an autonomous driving platform. We showcase new avenues for blending foundation models with 3D open-set multimodal mapping.

Operating a robot in the open world requires a high level of robustness with respect to previously unseen environments. Optimally, the robot is able to adapt by itself to new conditions without human supervision, e.g., automatically adjusting its perception system to changing lighting conditions. In this work, we address the task of continual learning for deep learning-based monocular depth estimation and panoptic segmentation in new environments in an online manner. We introduce CoDEPS to perform continual learning involving multiple real-world domains while mitigating catastrophic forgetting by leveraging experience replay. In particular, we propose a novel domain-mixing strategy to generate pseudo-labels to adapt panoptic segmentation. Furthermore, we explicitly address the limited storage capacity of robotic systems by leveraging sampling strategies for constructing a fixed-size replay buffer based on rare semantic class sampling and image diversity. We perform extensive evaluations of CoDEPS on various real-world datasets demonstrating that it successfully adapts to unseen environments without sacrificing performance on previous domains while achieving state-of-the-art results. The code of our work is publicly available at http://codeps.cs.uni-freiburg.de.

Modular robots hold the promise of changing their shape and even dimension to adapt to various tasks and environments. To realize this superiority, it is essential to find the appropriate morphology and its corresponding behavior simultaneously to ensure optimality of the reconfiguration. However, achieving co-optimization is challenging because robotic configuration and motion are interactive and coupled with each other, as well as their optimization processes. To this end, we proposed a co-optimization framework based on hierarchical Deep Reinforcement Learning (DRL), consisting of a configuration model and a motion model based on the Twin Delayed Deep Deterministic policy gradient algorithm (TD3). The two network models update asynchronously with a shared reward to ensure co-optimality. We conduct simulations and experiments with the Webots platform to validate the proposed framework, and the preliminary results show that it yields high quality optimization schemes and thus allows modular robots to be more adaptive to dynamic and multi-task scenarios.

Grasping small objects surrounded by unstable or non-rigid material plays a crucial role in applications such as surgery, harvesting, construction, disaster recovery, and assisted feeding. This task is especially difficult when fine manipulation is required in the presence of sensor noise and perception errors; errors inevitably trigger dynamic motion, which is challenging to model precisely. Circumventing the difficulty to build accurate models for contacts and dynamics, data-driven methods like reinforcement learning (RL) can optimize task performance via trial and error, reducing the need for accurate models of contacts and dynamics. Applying RL methods to real robots, however, has been hindered by factors such as prohibitively high sample complexity or the high training infrastructure cost for providing resets on hardware. This work presents CherryBot, an RL system that uses chopsticks for fine manipulation that surpasses human reactiveness for some dynamic grasping tasks. By integrating imprecise simulators, suboptimal demonstrations and external state estimation, we study how to make a real-world robot learning system sample efficient and general while reducing the human effort required for supervision. Our system shows continual improvement through 30 minutes of real-world interaction: through reactive retry, it achieves an almost 100% success rate on the demanding task of using chopsticks to grasp small objects swinging in the air. We demonstrate the reactiveness, robustness and generalizability of CherryBot to varying object shapes and dynamics (e.g., external disturbances like wind and human perturbations). Videos are available at https://goodcherrybot.github.io/.

In this work, we extend the model predictive control methods developed in the legged robotics literature to collaborative loco-manipulation settings. The systems we study entail a payload collectively carried by multiple quadruped robots equipped with a mechanical arm. We use a direct multiple shooting method to solve the resulting high-dimensional, optimal control problems for trajectories of ground reaction forces, manipulation wrenches, and stepping locations. To capture the dominant dynamics of the system, we model each agent and the shared payload as single rigid bodies. We demonstrate the versatility of our framework in a series of simulation experiments involving collaborative manipulation over challenging terrains.

Developing the next generation of household robot helpers requires combining locomotion and interaction capabilities, which is generally referred to as mobile manipulation (MoMa). MoMa tasks are difficult due to the large action space of the robot and the common multi-objective nature of the task, e.g., efficiently reaching a goal while avoiding obstacles. Current approaches often segregate tasks into navigation without manipulation and stationary manipulation without locomotion by manually matching parts of the action space to MoMa sub-objectives (e.g. base actions for locomotion objectives and arm actions for manipulation). This solution prevents simultaneous combinations of locomotion and interaction degrees of freedom and requires human domain knowledge for both partitioning the action space and matching the action parts to the sub-objectives. In this paper, we introduce Causal MoMa, a new framework to train policies for typical MoMa tasks that makes use of the most favorable subspace of the robot’s action space to address each sub-objective. Causal MoMa automatically discovers the causal dependencies between actions and terms of the reward function and exploits these dependencies in a causal policy learning procedure that reduces gradient variance compared to previous state-of-the-art policy gradient algorithms, improving convergence and results. We evaluate the performance of Causal MoMa on three types of simulated robots across different MoMa tasks and demonstrate success in transferring the policies trained in simulation directly to a real robot, where our agent is able to follow moving goals and react to dynamic obstacles while simultaneously and synergistically controlling the whole-body: base, arm, and head. More information at https://sites.google.com/view/causal-moma

We propose CLIP-Fields, an implicit scene model that can be used for a variety of tasks, such as segmentation, instance identification, semantic search over space, and view localization. CLIP-Fields learns a mapping from spatial locations to semantic embedding vectors. Importantly, we show that this mapping can be trained with supervision coming only from web-image and web-text trained models such as CLIP, Detic, and Sentence-BERT; and thus uses no direct human supervision. When compared to baselines like Mask-RCNN, our method outperforms on few-shot instance identification or semantic segmentation on the HM3D dataset with only a fraction of the examples. Finally, we show that using CLIP-Fields as a scene memory, robots can perform semantic navigation in real-world environments. Our code and demonstration videos are available here: https://clip-fields.github.io

This paper proposes a novel method for estimating the set of plausible poses of a rigid object from a set of points with volumetric information, such as whether each point is in free space or on the surface of the object. In particular, we study how pose can be estimated from force and tactile data arising from contact. Using data derived from contact is challenging because it is inherently less information-dense than visual data, and thus the pose estimation problem is severely under-constrained when there are few contacts. Rather than attempting to estimate the true pose of the object, which is not tractable without a large number of contacts, we seek to estimate a plausible set of poses which obey the constraints imposed by the sensor data. Existing methods struggle to estimate this set because they are either designed for single pose estimates or require informative priors to be effective. Our approach to this problem, Constrained pose Hypothesis Set Elimination (CHSEL), has three key attributes: 1) It considers volumetric information, which allows us to account for known free space; 2) It uses a novel differentiable volumetric cost function to take advantage of powerful gradient-based optimization tools; and 3) It uses methods from the Quality Diversity (QD) optimization literature to produce a diverse set of high-quality poses. To our knowledge, QD methods have not been used previously for pose registration. We also show how to update our plausible pose estimates online as more data is gathered by the robot. Our experiments suggest that CHSEL shows large performance improvements over several baseline methods for both simulated and real-world data.

Imitation learning holds great promise for addressing the complex task of autonomous urban driving, as experienced human drivers can navigate highly challenging scenarios with ease. While behavior cloning is a widely used imitation learning approach in autonomous driving due to its exemption from risky online interactions, it suffers from the covariate shift issue. To address this limitation, we propose a context-conditioned imitation learning approach that employs a policy to map the context state into the ego vehicle's future trajectory, rather than relying on the traditional formulation of both ego and context states to predict the ego action. Additionally, to reduce the implicit ego information in the coordinate system, we design an ego-perturbed goal-oriented coordinate system. The origin of this coordinate system is the ego vehicle's position plus a zero mean Gaussian perturbation, and the x-axis direction points towards its goal position. Our experiments on the real-world large-scale Lyft and nuPlan datasets show that our method significantly outperforms state-of-the-art approaches.

In recent years, learning-based control in robotics has gained significant attention due to its capability to address complex tasks in real-world environments. With the advances in machine learning algorithms and computational capabilities, this approach is becoming increasingly important for solving challenging control problems in robotics by learning unknown or partially known robot dynamics. Active exploration, in which a robot directs itself to states that yield the highest information gain, is essential for efficient data collection and minimizing human supervision. Similarly, uncertainty-aware deployment has been a growing concern in robotic control, as uncertain actions informed by the learned model can lead to unstable motions or failure. However, active exploration and uncertainty-aware deployment have been studied independently, and there is limited literature that seamlessly integrates them. This paper presents a unified model-based reinforcement learning framework that bridges these two tasks in the robotics control domain. Our framework uses a probabilistic ensemble neural network for dynamics learning, allowing the quantification of epistemic uncertainty via Jensen-Rényi Divergence. The two opposing tasks of exploration and deployment are optimized through state-of-the-art sampling-based MPC, resulting in efficient collection of training data and successful avoidance of uncertain state-action spaces. We conduct experiments on both autonomous vehicles and wheeled robots, showing promising results for both exploration and deployment.