IROS 2023 Papers

A key aspect of driving a road vehicle is to interact with other road users, assess their intentions and make riskaware tactical decisions. An intuitive approach to enabling an intelligent automated driving system would be incorporating some aspects of human driving behavior. To this end, we propose a novel driving framework for egocentric views based on spatio-temporal traffic graphs. The traffic graphs model not only the spatial interactions amongst the road users but also their individual intentions through temporally associated message passing. We leverage a spatio-temporal graph convolutional network (ST-GCN) to train the graph edges. These edges are formulated using parameterized functions of 3D positions and scene-aware appearance features of road agents. Along with tactical behavior prediction, it is crucial to evaluate the risk-assessing ability of the proposed framework. We claim that our framework learns risk-aware representations by improving on the task of risk object identification, especially in identifying objects with vulnerable interactions like pedestrians and cyclists.

Camera localization is a fundamental problem for many applications in computer vision, robotics, and autonomy. Despite recent deep learning-based approaches, the lack of robustness in challenging conditions persists due to changes in appearance caused by texture-less planes, repeating structures, reflective surfaces, motion blur, and illumination changes. Data augmentation is an attractive solution, but standard image perturbation methods fail to improve localization robustness. To address this, we propose RADA, which concentrates on perturbing the most vulnerable pixels to generate relatively less image perturbations that perplex the network. Our method outperforms previous augmentation techniques, achieving up to twice the accuracy of state-of-the-art models even under ‘unseen’ challenging weather conditions. Videos of our results can be found at https://youtu.be/niOv7-fJeCA. The source code for RADA is publicly available at https://github.com/jialuwang123321/RADA.

This paper describes RACECAR, the first open dataset for full-scale and high-speed autonomous racing. Multi-modal sensor data was collected from fully autonomous Indy race cars operating at speeds of up to 170 mph (273 kph). Six teams who raced in the Indy Autonomous Challenge (2021–2022) have contributed to this dataset. The dataset spans 11 racing scenarios across two race tracks which include solo laps, multi-agent laps, overtaking situations, high-accelerations, banked tracks, obstacle avoidance, pit entry and exit at different speeds. The dataset contains 27 racing sessions across 11 scenarios with over 6.5 hours of autonomous racing. The data has been released in both ROS 2 and nuScenes format. We have also developed the ROSbag2nuScenes conversion library to achieve this. The RACECAR data is unique because of the high-speed environment of autonomous racing. We present several benchmark problems on localization, object detection and tracking (LiDAR, Radar, and Camera), and mapping to explore issues that arise at the limits of operation of the vehicle. RACECAR data can be accessed at https://github.com/linklab-uva/RACECAR_DATA.

Deep learning-based models are at the top of most driver observation benchmarks due to their remarkable accuracies but come with a high computational cost, while the resources are often limited in real-world driving scenarios. This paper presents a lightweight framework for resource- efficient driver activity recognition. We enhance 3D MobileNet, a speed-optimized neural architecture for video classification, with two paradigms for improving the trade-off between model accuracy and computational efficiency: knowledge distillation and model quantization. Knowledge distillation prevents large drops in accuracy when reducing the model size by harvesting knowledge from a large teacher model (I3D) via soft labels instead of using the original ground truth. Quantization further drastically reduces the memory and computation requirements by representing the model weights and activations using lower precision integers. Extensive experiments on a public dataset for in-vehicle monitoring during autonomous driving show that our proposed framework leads to an 3- fold reduction in model size and 1.4-fold improvement in inference time compared to an already speed-optimized architecture. Our code is available at https://github.com/calvintanama/qd-driver-activity-reco.

Legged robots are increasingly entering new domains and applications, including search and rescue, inspection, and logistics. However, for such a systems to be valuable in real-world scenarios, they must be able to autonomously and robustly navigate irregular terrains. In many cases, robots that are sold on the market do not provide such abilities, being able to perform only blind locomotion. Furthermore, their controller cannot be easily modified by the end-user, requiring a new and time-consuming control synthesis. In this work, we present a fast local motion planning pipeline that extends the capabilities of a black-box walking controller that is only able to track high-level reference velocities. More precisely, we learn a set of motion models for such a controller that maps high-level velocity commands to Center of Mass (CoM) and footstep motions. We then integrate these models with a variant of the AA * algorithm to plan the CoM trajectory, footstep sequences, and corresponding high-level velocity commands based on visual information, allowing the quadruped to safely traverse irregular terrains at demand.

Robotic cloth manipulation is an increasingly relevant area of research, challenging classic control algorithms due to the deformable nature of cloth. While it is possible to apply linear model predictive control to make the robot move the cloth according to a given reference, this approach suffers from a large dimensionality of the state-space representation of the cloth models. To address this issue, in this work we study the application of an input-output model predictive control strategy, based on quadratic dynamic matrix control, to robotic cloth manipulation. To account for uncertain disturbances on the cloth's motion, we further extend the algorithm with suitable chance constraints. In extensive simulated experiments, involving disturbances and obstacle avoidance, we show that quadratic dynamic matrix control can be successfully applied in different cloth manipulation scenarios, with significant gains in optimization speed compared to standard model predictive control strategies. The experiments further demonstrate that the closed-loop model used by quadratic dynamic matrix control can be beneficial to the tracking accuracy, leading to improvements over the standard predictive control strategy. Moreover, a preliminary experiment on a real robot shows that quadratic dynamic matrix control can indeed be employed in real settings.

Pre-defined manipulation primitives are widely used for cloth manipulation. However, cloth properties such as its stiffness or density can highly impact the performance of these primitives. Although existing solutions have tackled the parameterisation of pick and place locations, the effect of factors such as the velocity or trajectory of quasi-static and dynamic manipulation primitives has been neglected. Choosing appropriate values for these parameters is crucial to cope with the range of materials present in house-hold cloth objects. To address this challenge, we introduce the Quasi-Dynamic Parameterisable (QDP) method, which optimises parameters such as the motion velocity in addition to the pick and place positions of quasi-static and dynamic manipulation primitives. In this work, we leverage the framework of Sequential Reinforcement Learning to decouple sequentially the parameters that compose the primitives. To evaluate the effectiveness of the method, we focus on the task of cloth unfolding with a robotic arm in simulation and real-world experiments. Our results in simulation show that by deciding the optimal parameters for the primitives the performance can improve by 20% compared to sub-optimal ones. Real-world results demonstrate the advantage of modifying the velocity and height of manipulation primitives for cloths with different mass, stiffness, shape, and size. Supplementary material, videos, and code, can be found at https://sites.google.com/view/qdp-srl.

Global point cloud registration is essential in many robotics tasks like loop closing and relocalization. Unfortunately, the registration often suffers from the low overlap between point clouds, a frequent occurrence in practical applications due to occlusion and viewpoint change. In this paper, we propose a graph-theoretic framework to address the problem of global point cloud registration with low overlap. To this end, we construct a consistency graph to facilitate robust data association and employ graduated non-convexity (GNC) for reliable pose estimation, following the state-of-the-art (SoTA) methods. Unlike previous approaches, we use semantic cues to scale down the dense point clouds, thus reducing the problem size. Moreover, we address the ambiguity arising from the consistency threshold by constructing a pyramid graph with multi-level consistency thresholds. Then we propose a cascaded gradient ascend method to solve the resulting densest clique problem and obtain multiple pose candidates for every consistency threshold. Finally, fast geometric verification is employed to select the optimal estimation from multiple pose candidates. Our experiments, conducted on a self-collected indoor dataset and the public KITTI dataset, demonstrate that our method achieves the highest success rate despite the low overlap of point clouds and low semantic quality. We have open-sourced our code1 for this project.

For robotic systems to interact with objects in dynamic environments, it is essential to perceive the physical properties of the objects such as shape, friction coefficient, mass, center of mass, and inertia. This not only eases selecting manipulation action but also ensures the task is performed as desired. However, estimating the physical properties of especially novel objects is a challenging problem, using either vision or tactile sensing. In this work, we propose a novel framework to estimate key object parameters using non-prehensile manipulation using vision and tactile sensing. Our proposed active dual differentiable filtering (ADDF) approach as part of our framework learns the object-robot interaction during non-prehensile object push to infer the object's parameters. Our proposed method enables the robotic system to employ vision and tactile information to interactively explore a novel object via non-prehensile object push. The novel proposed NN-step active formulation within the differentiable filtering facilitates efficient learning of the object-robot interaction model and during inference by selecting the next best exploratory push actions (where to push? and how to push?). We extensively evaluated our framework in simulation and real-robotic scenarios, yielding superior performance to the state-of-the-art baseline.

We focus on the problem of rearranging a set of objects with a team of car-like robot pushers built using off-the-shelf components. Maintaining control of pushed objects while avoiding collisions in a tight space demands highly coordinated motion that is challenging to execute on constrained hardware. Centralized replanning approaches become intractable even for small-sized problems whereas decentralized approaches often get stuck in deadlocks. Our key insight is that by carefully assigning pushing tasks to robots, we could reduce the complexity of the rearrangement task, enabling robust performance via scalable decentralized control. Based on this insight, we built PuSHR, a system that optimally assigns pushing tasks and trajectories to robots offline, and performs trajectory tracking via decentralized control online. Through an ablation study in simulation, we demonstrate that PuSHR dominates baselines ranging from purely centralized to fully decentralized in terms of success rate and time efficiency across challenging tasks with up to 4 robots. Hardware experiments demonstrate the transfer of our system to the real world and highlight its robustness to model inaccuracies. Our code can be found at https://github.com/prl-mushr/pushr, and videos from our experiments at https://youtu.be/nyUn9mHoR8Y.

Recently, the feature-level generation has demonstrated the effectiveness of pseudo-stereo synthesis in Monocular 3D Detection (M3D). In this paper, we aim to further bridge the gap between the stereo and the monocular 3D object detectors in autonomous driving through direct image-level pseudo-stereo generation. We propose a novel Cycled Generative Pseudo-Stereo (CGPS) architecture to generate the right-view image from the left-view for constructing a pseudo-stereo pair to stereo 3D object detectors while maintaining the natural of M3D with the left-view image as the only input. Moreover, we use a triplet consistency loss to focus on the detected objects in the pseudo-stereo generation. Besides, we demonstrate that the proposed CGPS is an ad-hoc module to adapt top stereo 3D object detectors into monocular 3D object detectors. The proposed framework with CGPS achieves 74.80%, 55.28%, and 46.70% 3DAP for easy, moderate, and hard difficulty levels in monocular 3D detection on the KITTI benchmark with comparable performance to the stereo 3D object detectors but using a monocular image as the only input. Till the submission, the proposed M3D framework ranks 1st with dramatic improvements against the existing monocular 3D detectors on the KITTI benchmark.

Robots reason about the environment through dedicated representations. Despite the fact that Gaussian Process (GP)-based representations are appealing due to their probabilistic and continuous nature, the cubic computational complexity is a concern. In this paper, we present a novel efficient GP-based representation that has the ability to produce accurate distance fields and is parameterised by the optimal locations of pseudo inputs. When applying the proposed method together with a kernel approximation approach, we show it outperforms well-established sparse GP frameworks in efficiency and accuracy. Moreover, we extend the proposed method to work in a dynamic setting, where a map is built iteratively and the scene dynamics are accounted for by adding or removing objects from the environment representation. In a nutshell, our method provides the ability to infer dynamic distance fields and achieve state-of-the-art reconstruction efficiently.

Planning a path for a mobile robot typically requires building a map (e.g., an occupancy grid) of the environment as the robot moves around. While navigating in an unknown environment, the map built by the robot online may have many as-yet-unknown regions. A conservative planner may avoid such regions taking a longer time to navigate to the goal. Instead, if a robot is able to correctly predict the occupancy in the occluded regions, the robot may navigate efficiently. We present a self-supervised occupancy prediction technique, ProxMaP, to predict the occupancy within the proximity of the robot to enable faster navigation. We show that ProxMaP generalizes well across realistic and real domains, and improves the robot navigation efficiency in simulation by 12.40% against a traditional navigation method. We share our findings and code at https://raaslab.org/projects/ProxMaP.

This paper develops a provably stable sensor-driven controller for path-following applications of robots with unicycle kinematics, one specific class of which is the wheeled mobile robot (WMR). The sensor measurement is converted to a scalar value (the score) through some mapping (the score function); the latter may be designed or learned. The score is then mapped to forward and angular velocities using a simple rule with three parameters. The key contribution is that the correctness of this controller only relies on the score function satisfying monotonicity conditions with respect to the underlying state - local path coordinates - instead of achieving specific values at all states. The monotonicity conditions may be checked online by moving the WMR, without state estimation, or offline using a generative model of measurements such as in a simulator. Our approach provides both the practicality of a purely measurement-based control and the correctness of state-based guarantees. We demonstrate the effectiveness of this path-following approach on both a simulated and a physical WMR that use a learned score function derived from a binary classifier trained on real depth images.

Although domestic service robots are expected to assist individuals who require support, they cannot currently interact smoothly with people through natural language. For example, given the instruction “Bring me a bottle from the kitchen,” it is difficult for such robots to specify the bottle in an indoor environment. Most conventional models have been trained on real-world datasets that are labor-intensive to collect, and they have not fully leveraged simulation data through a transfer learning framework. In this study, we propose a novel transfer learning approach for multimodal language understanding called Prototypical Contrastive Transfer Learning (PCTL), which uses a new contrastive loss called Dual ProtoNCE. We introduce PCTL to the task of identifying target objects in domestic environments according to free-form natural language instructions. To validate PCTL, we built new real-world and simulation datasets. Our experiment demonstrated that PCTL outperformed existing methods. Specifically, PCTL achieved an accuracy of 78.1 %, whereas simple fine-tuning achieved an accuracy of 73.4 %.

Soft actuators have several advantages, including large deformation, safety and adaptability to the environment, and shock absorbance. However, they are weak against sharp objects owing to their soft bodies. This paper proposes a novel protective skin mechanism with an exhaustive arrangement of tiny rigid bodies. Small pieces were sewed on an elastic sheet using Kevlar strings. We conducted some measurements of puncture resistance, elasticity, and descaling resistance of the scale mechanisms. The results indicate that the elasticity of the proposed scale mechanism was just 150% larger than a simple silicone sheet. In addition, approximately 15 N was required for descaling, which is seven times larger than that of the glued scales. There was no puncture even when pricked with a needle.

The estimation of external joint torque and contact wrench is essential for achieving stable locomotion of humanoids and safety-oriented robots. Although the contact wrench on the foot of humanoids can be measured using a force-torque sensor (FTS), FTS increases the cost, inertia, complexity, and failure possibility of the system. This paper introduces a method for learning external joint torque solely using proprioceptive sensors (encoders and IMUs) for a floating base robot. For learning, the GRU network is used and random walking data is collected. Real robot experiments demonstrate that the network can estimate the external torque and contact wrench with significantly smaller errors compared to the model-based method, momentum observer (MOB) with friction modeling. The study also validates that the estimated contact wrench can be utilized for zero moment point (ZMP) feedback control, enabling stable walking. Moreover, even when the robot's feet and the inertia of the upper body are changed, the trained network shows consistent performance with a model-based calibration. This result demonstrates the possibility of removing FTS on the robot, which reduces the disadvantages of hardware sensors.

Legged robots leverage ground contacts and the reaction forces they provide to achieve agile locomotion. However, uncertainty coupled with contact discontinuities can lead to failure, especially in real-world environments with unexpected height variations such as rocky hills or curbs. To enable dynamic traversal of extreme terrain, this work introduces 1) a proprioception-based gait planner for estimating unknown hybrid events due to elevation changes and responding by modifying contact schedules and planned footholds online, and 2) a two-degree-of-freedom tail for improving contact-independent control and a corresponding decoupled control scheme for better versatility and efficiency. Simulation results show that the gait planner significantly improves stability under unforeseen terrain height changes compared to methods that assume fixed contact schedules and footholds. Further, tests have shown that the tail is particularly effective at maintaining stability when encountering a terrain change with an initial angular disturbance. The results show that these approaches work synergistically to stabilize locomotion with elevation changes up to 1.5 times the leg length and tilted initial states.

Entanglements like vines and branches in natural settings or cords and pipes in human spaces prevent mobile robots from accessing many environments. Legged robots should be effective in these settings, and more so than wheeled or tracked platforms, but naive controllers quickly become entangled and stuck. In this paper we present a method for proprioception aimed specifically at the task of sensing entanglements of a robot's legs as well as a reaction strategy to disentangle legs during their swing phase as they advance to their next foothold. We demonstrate our proprioception and reaction strategy enables traversal of entanglements of many stiffnesses and geometries succeeding in 14 out of 16 trials in laboratory tests, as well as a natural outdoor environment.