IROS 2024 论文列表

Model-free reinforcement learning is a promising approach for autonomously solving challenging robotics control problems, but faces exploration difficulty without information about the robot’s morphology. The under-exploration of multiple modalities with symmetric states leads to behaviors that are often unnatural and sub-optimal. This issue becomes particularly pronounced in the context of robotic systems with morphological symmetries, such as legged robots for which the resulting asymmetric and aperiodic behaviors compromise performance, robustness, and transferability to real hardware. To mitigate this challenge, we can leverage symmetry to guide and improve the exploration in policy learning via equivariance / invariance constraints. We investigate the efficacy of two approaches to incorporate symmetry: modifying the network architectures to be strictly equivariant / invariant, and leveraging data augmentation to approximate equivariant / invariant actor-critics. We implement the methods on challenging loco-manipulation and bipedal locomotion tasks and compare with an unconstrained baseline. We find that the strictly equivariant policy consistently outperforms other methods in sample efficiency and task performance in simulation. Additionaly, symmetry-incorporated approaches exhibit better gait quality, higher robustness and can be deployed zero-shot to hardware.

Optimizing robotic action parameters is a significant challenge for manipulation tasks that demand high levels of precision and generalization. Using a model-based approach, the robot must quickly reason about the outcomes of different actions using a predictive model to find a set of parameters that will have the desired effect. The model may need to capture the behaviors of rigid and deformable objects, as well as objects of various shapes and sizes. Predictive models often need to trade-off speed for prediction accuracy and generalization. This paper proposes a framework that leverages the strengths of multiple predictive models, including analytical, learned, and simulation-based models, to enhance the efficiency and accuracy of action parameter optimization. Our approach uses Model Deviation Estimators (MDEs) to determine the most suitable predictive model for any given state-action parameters, allowing the robot to select models to make fast and precise predictions. We extend the MDE framework by not only learning sim-to-real MDEs, but also sim-to-sim MDEs. Our experiments show that these sim-to-sim MDEs provide significantly faster parameter optimization as well as a basis for efficiently learning sim-to-real MDEs through finetuning. The ease of collecting sim-to-sim training data also allows the robot to learn MDEs based directly on visual inputs and local material properties.

Classical structural-based visual localization methods offer high accuracy but face trade-offs in terms of storage, speed, and privacy. A recent innovation, keypoint scene coordinate regression (KSCR) named D2S addresses these issues by leveraging graph attention networks to enhance keypoint relationships and predict their 3D coordinates using a simple multilayer perceptron (MLP). Camera pose is then determined via PnP+RANSAC, using established 2D-3D correspondences. While KSCR achieves competitive results, rivaling state-of-the-art image-retrieval methods like HLoc across multiple benchmarks, its performance is hindered when data samples are limited due to the deep learning model’s reliance on extensive data. This paper proposes a solution to this challenge by introducing a pipeline for keypoint descriptor synthesis using Neural Radiance Field (NeRF). By generating novel poses and feeding them into a trained NeRF model to create new views, our approach enhances the KSCR’s generalization capabilities in data-scarce environments. The proposed system could significantly improve localization accuracy by up to 50% and cost only a fraction of time for data synthesis. Furthermore, its modular design allows for the integration of multiple NeRFs, offering a versatile and efficient solution for visual localization. The implementation is publicly available at: https://github.com/ais-lab/DescriptorSynthesis4Feat2Map.

Maps are essential for diverse applications, such as vehicle navigation and autonomous robotics. Both require spatial models for effective route planning and localization. This paper addresses the challenge of road graph construction for autonomous vehicles. Despite recent advances, creating a road graph remains labor-intensive and has yet to achieve full automation. The goal of this paper is to generate such graphs automatically and accurately. Modern cars are equipped with onboard sensors used for today's advanced driver assistance systems like lane keeping. We propose using global navigation satellite system (GNSS) traces and basic image data acquired from these standard sensors in consumer vehicles to estimate road-level maps with minimal effort. We exploit the spatial information in the data by framing the problem as a road centerline semantic segmentation task using a convolutional neural network. We also utilize the data’s time series nature to refine the neural network’s output by using map matching. We implemented and evaluated our method using a fleet of real consumer vehicles, only using the deployed onboard sensors. Our evaluation demonstrates that our approach not only matches existing methods on simpler road configurations but also significantly outperforms them on more complex road geometries and topologies. This work received the 2023 Woven by Toyota Invention Award.

This article studies the commonsense object affordance concept for enabling close-to-human task planning and task optimization of embodied robotic agents in urban environments. The focus of the object affordance is on reasoning how to effectively identify object’s inherent utility during the task execution, which in this work is enabled through the analysis of contextual relations of sparse information of 3D scene graphs. The proposed framework develops a Correlation Information (CECI) model to learn probability distributions using a Graph Convolutional Network, allowing to extract the commonsense affordance for individual members of a semantic class. The overall framework was experimentally validated in a real-world indoor environment, showcasing the ability of the method to level with human commonsense. For a video of the article, showcasing the experimental demonstration, please refer to the following link: https://youtu.be/BDCMVx2GiQE

Autonomous offroad driving is essential for applications like emergency rescue, military operations, and agriculture. Despite progress, systems struggle with high-speed vehicles exceeding 10m/s due to the need for accurate long-range (> 50m) perception for safe navigation. Current approaches are limited by sensor constraints; LiDAR-based methods offer precise short-range data but are noisy beyond 30m, while visual models provide dense long-range measurements but falter with unseen scenarios. To overcome these issues, we introduce ALTER, a learning-on-the-drive perception framework that leverages both sensor types. ALTER uses a self-supervised visual model to learn and adapt from near-range LiDAR measurements, improving long-range prediction in new environments without manual labeling. It also includes a model selection module for better sensor failure response and adaptability to known environments. Testing in two real-world settings showed on average 43.4% better traversability prediction than LiDAR-only and 164% over non-adaptive state-of-the-art (SOTA) visual semantic methods after 45 seconds of online learning.

We present a novel approach to learning-augmented, long-horizon navigation under uncertainty in large-scale environments in which considering the robot dynamics is essential for informing good behavior. Our approach tightly integrates sampling-based motion planning, which computes dynamically feasible routes to the goal through different unexplored boundaries, and a high-level planner that leverages predictions about unseen space to select a route that best makes progress toward the unseen goal. Owing to its ability to understand the impacts of the robot’s dynamics on how it should attempt to reach the goal, our approach achieves both higher reliability and improved navigation performance compared to competitive learning-informed and non-learned baselines in simulated office-building-like environments.

A quadruped robot faces balancing challenges on a six-degrees-of-freedom moving platform, like subways, buses, airplanes, and yachts, due to independent platform motions and resultant diverse inertia forces on the robot. To alleviate these challenges, we present the Learning-based Active Stabilization on Moving Platforms (LAS-MP), featuring a self-balancing policy and system state estimators. The policy adaptively adjusts the robot’s posture in response to the platform’s motion. The estimators infer robot and platform states based on proprioceptive sensor data. For a systematic training scheme across various platform motions, we introduce platform trajectory generation and scheduling methods. Our evaluation demonstrates superior balancing performance across multiple metrics compared to three baselines. Furthermore, we conduct a detailed analysis of the LAS-MP, including ablation studies and evaluation of the estimators, to validate the effectiveness of each component.

This paper proposes an adaptive admittance controller for improving efficiency and safety in physical human-robot interaction (pHRI) tasks in small-batch manufacturing that involve contact with stiff environments, such as drilling, polishing, cutting, etc. We aim to minimize human effort and task completion time while maximizing precision and stability during the contact of the machine tool attached to the robot’s end-effector with the workpiece. To this end, a two-layered learning-based human intention recognition mechanism is proposed, utilizing only the kinematic and kinetic data from the robot and two force sensors. A "subtask detector" recognizes the human intent by estimating which phase of the task is being performed, e.g., Idle, Tool-Attachment, Driving, and Contact. Simultaneously, a "motion estimator" continuously quantifies intent more precisely during the Driving to predict when Contact will begin. The controller is adapted online according to the subtask while allowing early adaptation before the Contact to maximize precision and safety and prevent potential instabilities. Three sets of pHRI experiments were performed with multiple subjects under various conditions. Spring compression experiments were performed in virtual environments to train the data-driven models and validate the proposed adaptive system, and drilling experiments were performed in the physical world to test the proposed methods’ efficacy in real-life scenarios. Experimental results show subtask classification accuracy of 84% and motion estimation R2 score of 0.96. Furthermore, 57% lower human effort was achieved during Driving as well as 53% lower oscillation amplitude at Contact as a result of the proposed system.

Robot multimodal locomotion encompasses the ability to transition between walking and flying, representing a significant challenge in robotics. This work presents an approach that enables automatic smooth transitions between legged and aerial locomotion. Leveraging the concept of Adversarial Motion Priors, our method allows the robot to imitate motion datasets and accomplish the desired task without the need for complex reward functions. The robot learns walking patterns from human-like gaits and aerial locomotion patterns from motions obtained using trajectory optimization. Through this process, the robot adapts the locomotion scheme based on environmental feedback using reinforcement learning, with the spontaneous emergence of mode-switching behavior. The results highlight the potential for achieving multimodal locomotion in aerial humanoid robotics through automatic control of walking and flying modes, paving the way for applications in diverse domains such as search and rescue, surveillance, and exploration missions. This research contributes to advancing the capabilities of aerial humanoid robots in terms of versatile locomotion in various environments. Video: https://youtu.be/mi6Do-x67CM

Automatically detecting and recovering from failures is an important but challenging problem for autonomous robots. Most of the recent work on learning to plan from demonstrations lacks the ability to detect and recover from errors in the absence of an explicit state representation and/or a (sub-) goal check function. We propose an approach (blending learning with symbolic search) for automated error discovery and recovery, without needing annotated data of failures. Central to our approach is a neuro-symbolic state representation, in the form of dense scene graph, structured based on the objects present within the environment. This enables efficient learning of the transition function and a discriminator that not only identifies failures but also localizes them facilitating fast re-planning via computation of heuristic distance function. We also present an anytime version of our algorithm, where instead of recovering to the last correct state, we search for a sub-goal in the original plan minimizing the total distance to the goal given a re-planning budget. Experiments on a physics simulator with a variety of simulated failures show the effectiveness of our approach compared to existing baselines, both in terms of efficiency as well as accuracy of our recovery mechanism.

Understanding collective behavior and how it evolves is important to ensure that robot swarms can be trusted in a shared environment. One way to understand the behavior of the swarm is through collective behavior reconstruction using prior demonstrations. Existing approaches often require access to the swarm controller which may not be available. We reconstruct collective behaviors in distinct swarm scenarios involving shared environments without using swarm controller information. We achieve this by transforming prior demonstrations into features that describe multi-agent interactions before behavior reconstruction with multi-agent generative adversarial imitation learning (MA-GAIL). We show that our approach outperforms existing algorithms in spatial organization, and can be used to observe and reconstruct a swarm’s behavior for further analysis and testing, which might be impractical or undesirable on the original robot swarm.

We consider the problem of training a fully convolutional network to estimate the relative 6D pose of a robot given a camera image, when the robot is equipped with independent controllable LEDs placed in different parts of its body. The training data is composed by few (or zero) images labeled with a ground truth relative pose and many images labeled only with the true state (ON or OFF) of each of the peer LEDs. The former data is expensive to acquire, requiring external infrastructure for tracking the two robots; the latter is cheap as it can be acquired by two unsupervised robots moving randomly and toggling their LEDs while sharing the true LED states via radio. Training with the latter dataset on estimating the LEDs’ state of the peer robot (pretext task) promotes learning the relative localization task (end task). Experiments on real-world data acquired by two autonomous wheeled robots show that a model trained only on the pretext task successfully learns to localize a peer robot on the image plane; fine-tuning such model on the end task with few labeled images yields statistically significant improvements in 6D relative pose estimation with respect to baselines that do not use pretext-task pre-training, and alternative approaches. Estimating the state of multiple independent LEDs promotes learning to estimate relative heading. The approach works even when a large fraction of training images do not include the peer robot and generalizes well to unseen environments.

Learning from Demonstration (LfD) is a promising approach to enable Multi-Robot Systems (MRS) to acquire complex skills and behaviors. However, the intricate interactions and coordination challenges in MRS pose significant hurdles for effective LfD. In this paper, we present a novel LfD framework specifically designed for MRS, which leverages visual demonstrations to capture and learn from robot-robot and robot-object interactions. Our framework introduces the concept of Interaction Keypoints (IKs) to transform the visual demonstrations into a representation that facilitates the inference of various skills necessary for the task. The robots then execute the task using sensorimotor actions and reinforcement learning (RL) policies when required. A key feature of our approach is the ability to handle unseen contact-based skills that emerge during the demonstration. In such cases, RL is employed to learn the skill using a classifier-based reward function, eliminating the need for manual reward engineering and ensuring adaptability to environmental changes. We evaluate our framework across a range of mobile robot tasks, covering both behavior-based and contact-based domains. The results demonstrate the effectiveness of our approach in enabling robots to learn complex multi-robot tasks and behaviors from visual demonstrations.

Velocity estimation is of great importance in autonomous racing. Still, existing solutions are characterized by limited accuracy, especially in the case of aggressive driving or poor generalization to unseen road conditions. To address these issues, we propose to utilize Unscented Kalman Filter (UKF) with a learned dynamics model that is optimized directly for the state estimation task. Moreover, we propose to aid this model with the online-estimated friction coefficient, which increases the estimation accuracy and enables zero-shot adaptation to the new road conditions. To evaluate the UKF-based velocity estimator with the proposed dynamics model, we introduced a publicly available dataset of aggressive maneuvers performed by an F1TENTH car, with side-slip angles reaching 40°. Using this dataset, we show that learning the dynamics model through UKF leads to improved estimation performance and that the proposed solution outperforms state-of-the-art learning-based state estimators by 17% in the nominal scenario. Moreover, we present unseen zero-shot adaptation abilities of the proposed method to the new road surface thanks to the proposed learning-based tire dynamics model with online friction estimation.

In this work, we consider the problem of learning end to end perception to control for ground vehicles solely from aerial imagery. Photogrammetric simulators allow the synthesis of novel views through the transformation of pre-generated assets into novel views. However, they have a large setup cost, require careful collection of data and often human effort to create usable simulators. We use a Neural Radiance Field (NeRF) as an intermediate representation to synthesize novel views from the point of view of a ground vehicle. These novel viewpoints can then be used for several downstream autonomous navigation applications. In this work, we demonstrate the utility of novel view synthesis though the application of training a policy for end to end learning from images and depth data. In a traditional real to sim to real framework, the collected data would be transformed into a visual simulator which could then be used to generate novel views. In contrast, using a NeRF allows a compact representation and the ability to optimize over the parameters of the visual simulator as more data is gathered in the environment. We demonstrate the efficacy of our method in a custom built mini-city environment through the deployment of imitation policies on robotic cars. We additionally consider the task of place localization and demonstrate that our method is able to relocalize the car in the real world.

Robotic grasping of a target object in cluttered environments poses considerable challenges, often due to limited collision-free grasp affordances caused by the close proximity of other objects. To overcome this limitation, non-prehensile actions like pushing can be strategically employed to manipulate the environment and improve the chances of successful grasps. In this paper, we introduce a novel pre-grasp manipulation policy designed to efficiently retrieve a target object from dense clutter by leveraging pushing actions and considering the gripper’s kinematic capabilities to strategically position the target object within the gripper’s closing region for a secure grasp. Unlike conventional approaches, our policy incorporates sequential pushing, allowing the robot to make decisions while within the camera’s field of view without retracting to a home position, leading to significantly reduced execution time per action. Our policy, trained in simulation, seamlessly transfers to real-world scenarios. Extensive experimental evaluation demonstrates superior performance, faster completion times, and robust generalization to unseen objects compared to existing baselines.

Actively guiding attention is an important mechanism to employ limited processing resources efficiently. The Recurrent Visual Attention Model (RAM) has been successfully applied to process large input images by sequentially attending to smaller image regions with an RL framework. In tactile perception, sequential attention methods are required naturally due to the limited size of the tactile receptive field. The concept of RAM was transferred to the haptic domain by the Haptic Attention Model (HAM) to iteratively generate a fixed number of informative haptic glances for tactile object classification. We extend HAM to a system capable of actively determining when sufficient haptic data is available for reliable classification. To this end, we introduce a hybrid action space, augmenting the continuous glance location with the discrete decision of when to classify. This allows balancing the cost of obtaining new samples against the cost of misclassification, resulting in an optimized number of glances while maintaining reasonable accuracy. We evaluate the efficiency of our approach on a handcrafted dataset, which allows us to compute the most efficient glance locations.

Quadruped robots are progressively being integrated into human environments. Despite the growing locomotion capabilities of quadrupedal robots, their interaction with objects in realistic scenes is still limited. While additional robotic arms on quadrupedal robots enable manipulating objects, they are sometimes redundant given that a quadruped robot is essentially a mobile unit equipped with four limbs, each possessing 3 degrees of freedom (DoFs). Hence, we aim to empower a quadruped robot to execute real-world manipulation tasks using only its legs. We decompose the loco-manipulation process into a low-level reinforcement learning (RL)-based controller and a high-level Behavior Cloning (BC)-based planner. By parameterizing the manipulation trajectory, we synchronize the efforts of the upper and lower layers, thereby leveraging the advantages of both RL and BC. Our approach is validated through simulations and real-world experiments, demonstrating the robot’s ability to perform tasks that demand mobility and high precision, such as lifting a basket from the ground while moving, closing a dishwasher, pressing a button, and pushing a door.