ICRA 2022 论文列表

Remote controlling robots without any automated help is difficult due to various limitations. Autocomplete mitigates this difficulty by automatically detecting and completing the intended motions on robots from the input of the user. Such an approach can improve the system performance and reduce the load on the operator. Usually, recognizing intended motions is achieved using pre-trained Deep Learning (DL) models. In this paper, we introduce personalization to the autocomplete teleoperation framework when new operators take over by customizing the autocomplete DL model using incremental learning. Also, we tackle the problem of concept drift that arises in real-life applications; the data distribution of already learned classes may change in unforeseen ways as new observations of these classes come sequentially over time. We create and update an exemplar set using new observations of the classes online so that the model can be trained to adapt to the new observations. Several scenarios have been evaluated to balance the speed of learning with the accuracy of the model, and results demonstrate the effectiveness of the proposed models and their advantage in adapting to the specific operator versus our previous framework: personalization using transfer learning with full feedback.

Incremental few-shot learning is highly expected for practical robotics applications. On one hand, robot is desired to learn new tasks quickly and flexibly using only few annotated training samples; on the other hand, such new additional tasks should be learned in a continuous and incremental manner without forgetting the previous learned knowledge dramatically. In this work, we propose a novel Class-Incremental Few- Shot Object Detection (CI-FSOD) framework that enables deep object detection network to perform effective continual learning from just few-shot samples without re-accessing the previous training data. We achieve this by equipping the widely-used Faster-RCNN detector with three elegant components. Firstly, to best preserve performance on the pre-trained base classes, we propose a novel Dual-Embedding-Space (DES) architecture which decouples the representation learning of base and novel categories into different spaces. Secondly, to mitigate the catastrophic forgetting on the accumulated novel classes, we propose a Sequential Model Fusion (SMF) method, which is able to achieve long-term memory without additional storage cost. Thirdly, to promote inter-task class separation in feature space, we propose a novel regularization technique that extends the classification boundary further away from the previous classes to avoid misclassification. Overall, our framework is simple yet effective and outperforms the previous SOTA with a significant margin of 2.4 points in AP performance.

Scene graphs represent the key components of a scene in a compact and semantically rich way, but are difficult to build during incremental SLAM operation because of the challenges of robustly identifying abstract scene elements and optimising continually changing, complex graphs. We present a distributed, graph-based SLAM framework for incrementally building scene graphs based on two novel components. First, we propose an incremental abstraction framework in which a neural network proposes abstract scene elements that are incorporated into the factor graph of a feature-based monocular SLAM system. Scene elements are confirmed or rejected through optimisation and incrementally replace the points yielding a more dense, semantic and compact representation. Second, enabled by our novel routing procedure, we use Gaussian Belief Propagation (GBP) for distributed inference on a graph processor. The time per iteration of GBP is structure-agnostic and we demonstrate the speed advantages over direct methods for inference of heterogeneous factor graphs. We run our system on real indoor datasets using planar abstractions and recover the major planes with significant compression.

Much of what we do as humans is engage socially with other agents, a skill that robots must also eventually possess. We demonstrate that a rich theory of social interactions originating from microsociology can be formalized by extending a nested MDP where agents reason about arbitrary functions of each other's rewards. This extended Social MDP allows us to encode the five basic interactions that underlie microsociology: cooperation, conflict, coercion, competition, and exchange. The result is a robotic agent capable of executing social interactions in new environments with no interaction-specific training; like humans it can engage socially in novel ways even without a single example of that social interaction. Moreover, the estimations of these Social MDPs align closely with the judge-ments of humans when considering which social interaction is taking place in an environment. This method both sheds light on the nature of social interactions, by providing concrete mathematical definitions, and brings rich social interactions into a mathematical framework that has proven to be natural for robotics.

In this paper we present a novel strategy for reactive collision-free feasible motion planning for robotic manipulators operating inside an environment populated by moving obstacles. The proposed strategy embeds the Dynamical System (DS) based obstacle avoidance algorithm into a constrained non-linear optimization problem following the Model Predictive Control (MPC) approach. The solution of the problem allows the robot to avoid undesired collision with moving obstacles ensuring at the same time that its motion is feasible and does not overcome the designed constraints on velocity and acceleration. Simulations demonstrate that the introduction of the MPC prediction horizon helps the optimization solver in finding the solution leading to obstacle avoidance in situations where a non predictive implementation of the DS-based method would fail. Finally, the proposed strategy has been validated in an experimental work-cell using a Franka-Emika Panda robot.

This paper presents the design and development of a robotic system to give physical assistance to the elderly or people with neurological disorders such as Ataxia or Parkin-son's. In particular, we propose using a mobile collaborative robot with an interaction-assistive whole-body interface to help people unable to maintain balance. The robotic system consists of an Omni-directional mobile base, a high-payload robotic arm, and an admittance-type interface acting as a support handle while measuring human-sourced interaction forces. The postural balance of the human body is estimated through the projection of the body Center of Mass (CoM) to the support polygon (SP) representing the quasi-static Center of Pressure (CoP). In response to the interaction forces and the tracking of the human posture, the robot can create assistive forces to restore balance in case of its loss. Otherwise, during normal stance or walking, it will follow the user with minimum/no opposing forces through the generation of coupled arm and base movements. As the balance-restoring strategy, we propose two strategies and evaluate them in a laboratory setting on healthy human participants. Quantitative and qualitative results of a 12-subjects experiment are then illustrated and discussed, comparing the performances of the two strategies and the overall system.

One of the key challenges to deep reinforcement learning (deep RL) is to ensure safety at both training and testing phases. In this work, we propose a novel technique of unsupervised action planning to improve the safety of on-policy reinforcement learning algorithms, such as trust region policy optimization (TRPO) or proximal policy optimization (PPO). We design our safety-aware reinforcement learning by storing all the history of “recovery” actions that rescue the agent from dangerous situations into a separate “safety” buffer and finding the best recovery action when the agent encounters similar states. Because this functionality requires the algorithm to query similar states, we implement the proposed safety mechanism using an unsupervised learning algorithm, k-means clustering. We evaluate the proposed algorithm on six robotic control tasks that cover navigation and manipulation. Our results show that the proposed safe RL algorithm can achieve higher rewards compared with multiple baselines in both discrete and continuous control problems. The supplemental video can be found at: https://youtu.be/AFTeWSohILo.

The shapes of most real-world objects are symmetric with respect to at least one plane of symmetry. This information is unconsciously used by humans when they attempt to estimate the shape of an object in presence of uncertainty or missing evidence, for example if the object is partially occluded or if they are exploring the object by touch (i.e. haptic exploration). In robotics, this concept has been used for the visual estimation of object shape. However, no attempt has been made so far to incorporate this idea into haptic-based estimation. This work presents a method for the haptic exploration of object shape that includes the assumption that a symmetry could exist. The approach combines a tailored version of Gaussian Processes and a novel exploratory procedure that is able to detect the position and orientation of any plane of symmetry. Our results show that, if one or more symmetries exist, the object shape can be estimated faster and more accurately. Interestingly, in the case that no symmetry is present, the exploration process is only slightly slower, and the final accuracy of the shape estimation is not compromised.

Human-Robot Collaboration (HRC) has become a major trend in robotics in recent years with the idea of combining the strengths from both humans and robots. In order to share the work to be done, many task planning approaches have been implemented. However, they don't fully satisfy the required adaptability in human-robot collaborative tasks, with most approaches not considering neither the state of the human partner nor the possibility of adapting the collaborative plan during execution or even anticipating failures. In this paper, we present a planning system for human-robot collaborative plans that takes into account the agents' states and deals with unforeseen human behaviour, by replanning in anticipation when the human state changes to prevent action failure. The human state is defined in terms of capacity, knowledge and motivation. The system has been implemented in a standardised environment using the Planning Domain Definition Language (PDDL) and the modular ROSPlan framework, and we have validated the approach in multiple simulation settings. Our results show that using the human model fosters an appropriate task allocation while allowing failure anticipation, replanning in time to prevent it.

We present a novel approach to improve 6 degree-of-freedom state propagation for unmanned aerial vehicles in a classical filter through pre-processing of high-speed inertial data with AI algorithms. We evaluate both an LSTM-based approach as well as a Transformer encoder architecture. Both algorithms take as input short sequences of fixed length N of high-rate inertial data provided by an inertial measurement unit (IMU) and are trained to predict in turn one pre-processed IMU sample that minimizes the state propagation error of a classical filter across M sequences. This setup allows us to provide sufficient temporal history to the networks for good performance while maintaining a high propagation rate of pre-processed IMU samples important for later deployment on real-world systems. In addition, our network architectures are formulated to directly accept input data at variable rates thus minimizing necessary data preprocessing. The results indicate that the LSTM based architecture outperforms the Transformer encoder architecture and significantly improves the propagation error even for long IMU propagation times.

Search-based techniques have shown great success in motion planning problems such as robotic navigation by discretizing the state space and precomputing motion primitives. However in domains with complex dynamic constraints, constructing motion primitives in a discretized state space is non-trivial. This requires operating in continuous space which can be challenging for search-based planners as they can get stuck in local minima regions. Previous work [1] on planning in continuous spaces introduced soft duplicate detection which requires search to compute the duplicity of a state with respect to previously seen states to avoid exploring states that are likely to be duplicates, especially in local minima regions. They propose a simple metric utilizing the Euclidean distance between states, and proximity to obstacles to compute the duplicity. In this paper, we improve upon this metric by introducing a kinodynamically informed metric, subtree overlap, between two states as the similarity between their successors that can be reached within a fixed time horizon using kinodynamic motion primitives. This captures the intuition that, due to robot dynamics, duplicate states can be far in Euclidean distance and result in very similar successor states, while non-duplicate states can be close and result in widely different successors. Our approach computes the new metric offline for a given robot dynamics, and stores the subtree overlap value for all possible relative state configurations. During search, the planner uses these precomputed values to speed up duplicity computation, and achieves fast planning times in continuous spaces in addition to completeness and sub-optimality guarantees. Empirically, we show that our improved metric for soft duplicity detection in search-based planning outperforms previous approaches in terms of planning time, by a factor of 1.5 to 2× on 3D and 5D planning domains with highly constrained dynamics.

Kalman-Particle Kernel Filter (KPKF) is a sub-class of Particle Filter (PF) that uses Gaussian kernels as particles, which enables a local Kalman update for each measurement in addition to the usual weight update. Besides, recent research about filtering on Lie groups brought powerful theoretical results, and showed the superiority of this approach. Hence, this paper extends the Euclidean KPKF to a new formulation on Lie groups and introduces substantial improvements based on Lie groups Kalman filters theory and Laplace Particle Filters on Lie groups (LG-LPF) for improved resampling. The proposed algorithm is tested on an angles-only UAV navigation scenario with challenging initial errors. It shows superior robustness and accuracy compared to Lie group Extended Kalman Filter (LG-EKF), with near-to optimal performance, even with a limited amount of particles.

Accurate identification of important objects in the scene is a prerequisite for safe and high-quality decision making and motion planning of intelligent agents (e.g., autonomous vehicles) that navigate in complex and dynamic environments. Most existing approaches attempt to employ attention mechanisms to learn importance weights associated with each object indirectly via various tasks (e.g., trajectory prediction), which do not enforce direct supervision on the importance estimation. In contrast, we tackle this task in an explicit way and formulate it as a binary classification (“important” or “unimportant”) problem. We propose a novel approach for important object identification in egocentric driving scenarios with relational reasoning on the objects in the scene. Besides, since human annotations are limited and expensive to obtain, we present a semi-supervised learning pipeline to enable the model to learn from unlimited unlabeled data. Moreover, we propose to leverage the auxiliary tasks of ego vehicle behavior prediction to further improve the accuracy of importance estimation. The proposed approach is evaluated on a public egocentric driving dataset (H3D) collected in complex traffic scenarios. A detailed ablative study is conducted to demonstrate the effectiveness of each model component and the training strategy. Our approach also outperforms rule-based baselines by a large margin.

Super-resolution of LiDAR range images is crucial to improving many downstream tasks such as object detection, recognition, and tracking. While deep learning has made a remarkable advances in super-resolution techniques, typical convolutional architectures limit upscaling factors to specific output resolutions in training. Recent work has shown that a continuous representation of an image and learning its implicit function enable almost limitless upscaling. However, the detailed approach, predicting values (depths) for neighbor pixels in the input and then linearly interpolating them, does not best fit the LiDAR range images since it does not fill the unmeasured details but creates a new image with regression in a high-dimensional space. In addition, the linear interpolation blurs sharp edges providing important boundary information of objects in 3-D points. To handle these problems, we propose a novel network, Implicit LiDAR Network (ILN), which learns not the values per pixels but weights in the interpolation so that the super-resolution can be done by blending the input pixel depths but with non-linear weights. Also, the weights can be considered as attentions from the query to the neighbor pixels, and thus an attention module in the recent Transformer architecture can be leveraged. Our experiments with a novel large-scale synthetic dataset demonstrate that the proposed network reconstructs more accurately than the state-of-the-art methods, achieving much faster convergence in training.

Action representation is an important yet often overlooked aspect in end-to-end robot learning with deep networks. Choosing one action space over another (e.g. target joint positions, or Cartesian end-effector poses) can result in surprisingly stark performance differences between various downstream tasks - and as a result, considerable research has been devoted to finding the right action space for a given application. However, in this work, we instead investigate how our models can discover and learn for themselves which action space to use. Leveraging recent work on implicit behavioral cloning, which takes both observations and actions as input, we demonstrate that it is possible to present the same action in multiple different spaces to the same policy - allowing it to learn inductive patterns from each space. Specifically, we study the benefits of combining Cartesian and joint action spaces in the context of learning manipulation skills. To this end, we present Implicit Kinematic Policies (IKP), which incorporates the kinematic chain as a differentiable module within the deep network. Quantitative experiments across several simulated continuous control tasks-from scooping piles of small objects, to lifting boxes with elbows, to precise block insertion with miscalibrated robots-suggest IKP not only learns complex prehensile and non-prehensile manipulation from pixels better than baseline alternatives, but also can learn to compensate for small joint encoder offset errors. Finally, we also run qualitative experiments on a real UR5e to demonstrate the feasibility of our algorithm on a physical robotic system with real data. See https://tinyurl.com/4wz3nf86 for code and supplementary material.

Over the past decade, the Differential Dynamic Programming (DDP) method has gained in maturity and popularity within the robotics community. Several recent contributions have led to the integration of constraints within the original DDP formulation, hence enlarging its domain of application while making it a strong and easy-to-implement competitor against alternative methods of the state of the art such as collocation or multiple-shooting approaches. Yet, and similarly to its competitors, DDP remains unable to cope with high-dimensional dynamics within a receding horizon fashion, such as in the case of online generation of athletic motion on humanoid robots. In this paper, we propose to make a step towards this objective by reformulating classical DDP as an implicit optimal control problem, allowing the use of more advanced integration schemes such as implicit or variational integrators. To that end, we introduce a primal-dual proximal Lagrangian approach capable of handling dynamical and path constraints in a unified manner, while taking advantage of the time sparsity inherent to optimal control problems. We show that this reformulation enables us to relax the dynamics along the optimization process by solving it inexactly: far from the optimality conditions, the dynamics are only partially fulfilled, but continuously enforced as the solver gets closer to the local optimal solution. This inexactness enables our approach to robustly handle large time steps (100 ms or more), unlike other DDP solvers of the state of the art, as experimentally validated through different robotic scenarii.

Impacts and other non-smooth behaviors are usually unwanted in robotic applications. However, several industrial tasks such as deburring, removing excess material, and assembling/fitting, involve impacts between objects, which can benefit from robotic automation due to the risks posed to human health. Towards this objective, in this paper, we propose a method for optimal impact planning and pre-configuration for torque-controlled robots. We thus employ a well-known impulsive contact model to plan the impact force and create a hierarchical quadratic programming based controller capable of minimizing the robot's peak torques by reconfiguring its joints optimally, before the impact occurs. The results obtained from multiple experiments during an industrial deburring task are discussed. Using a 7-DoF manipulator, we show consistent results, both in terms of accuracy of the impact force tracking with respect to the desired forces, and in terms of peak torques reduction and uniform torques distribution.

The ongoing COVID-19 pandemic has enforced governments across the world to impose social restrictions on the movement of people and confined them to their homes to avoid the spread of the disease. This not only forbids them from leaving their homes but also greatly reduces their physical activities. This situation has brought attention to virtual technologies such as virtual tours or telepresence robots. While these technologies allow people to remotely participate in activities, it does not address the problem of reduction in physical activities due to the pandemic. In this paper, we propose a telepresence robotic system driven by the user's gait to provide an immersive virtual walking experience in remote locations. To this end, we developed a control interface consisting of an automated treadmill that adjusts its speed to the user's pace automatically. This interface is used to control an avatar robot that sends a 360-degree live image back to the user for visual feedback. We conducted an evaluation experiment to compare the experience using the proposed system in two different conditions to that of regular walking. The results indicated that the proposed system gives an immersive and realistic virtual walking experience while demanding physical effort from the user.

Due to high complexity and occlusion, insufficient perception in the crowded urban intersection can be a serious safety risk for both human drivers and autonomous algorithms, whereas CVIS (Cooperative Vehicle Infrastructure System) is a proposed solution for full-participants perception under this scenario. However, the research on roadside multi-modal perception is still in its infancy, and there is no open-source data sets for such scene. Accordingly, this paper fills the gap. Through an IPS (Intersection Perception System) installed at the diagonal of the intersection, this paper proposes a high-quality multi-modal data sets for the intersection perception task. The center of the experimental intersection covers an area of 3000m2, and the extended distance reaches 300m, which is typical for CVIS. The first batch of open-source data includes 14198 frames, and each frame has an average of 319.84 labels, which is 9.6 times larger than the most crowded data sets (H3D data sets in 2019) by now. Our data sets is available at: http://www.openmpd.com/column/IPS300.