ICRA 2022 论文列表

Autonomous vehicles must reason about spatial occlusions in urban environments to ensure safety without being overly cautious. Prior work explored occlusion inference from observed social behaviors of road agents, hence treating people as sensors. Inferring occupancy from agent behaviors is an inherently multimodal problem; a driver may behave similarly for different occupancy patterns ahead of them (e.g., a driver may move at constant speed in traffic or on an open road). Past work, however, does not account for this multimodality, thus neglecting to model this source of aleatoric uncertainty in the relationship between driver behaviors and their environment. We propose an occlusion inference method that characterizes observed behaviors of human agents as sensor measurements, and fuses them with those from a standard sensor suite. To capture the aleatoric uncertainty, we train a conditional variational autoencoder with a discrete latent space to learn a multimodal mapping from observed driver trajectories to an occupancy grid representation of the view ahead of the driver. Our method handles multi-agent scenarios, combining measurements from multiple observed drivers using evidential theory to solve the sensor fusion problem. Our approach is validated on a cluttered, real-world intersection, outperforming baselines and demonstrating real-time capable performance. Our code is available at https://github.com/sisl/MultiAgentVariationalOcclusionInferenc

Multi-agent pathfinding (MAPF) has been widely used to solve large-scale real-world problems, e.g., automation warehouses. The learning-based, fully decentralized framework has been introduced to alleviate real-time problems and simultaneously pursue optimal planning policy. However, existing methods might generate significantly more vertex conflicts (or collisions), which lead to a low success rate or more makespan. In this paper, we propose a PrIoritized COmmunication learning method (PICO), which incorporates the implicit planning priorities into the communication topology within the decentralized multi-agent reinforcement learning framework. Assembling with the classic coupled planners, the implicit priority learning module can be utilized to form the dynamic communication topology, which also builds an effective collision-avoiding mechanism. PICO performs significantly better in large-scale MAPF tasks in success rates and collision rates than state-of-the-art learning-based planners.

The long-term goal of this work is to enable agents with low-information sensors to perform tasks usually restricted to ones with more sophisticated, high-information sensing capabilities. Our approach is to regulate the motion of these low-information agents to obtain “high-information” results. As a first step, we consider a multi-agent system tasked with locating and tracking a moving target using only noisy binary sensors that measure the presence (or lack thereof) of a target in the sensor's field of view. To generate effective paths for these agents, we use ergodic trajectory optimization with a novel mutual information map that is fast to compute and can handle the discontinuous measurement models often associated with low-information sensing. We compare our approach with existing motion planning methods in multiple simulated experiments. Our experiments show that agents using our method outperform purely coverage-based approaches as well as naive ergodic approaches.

We present a novel self-supervised algorithm named MotionHint for monocular visual odometry (VO) that takes motion constraints into account. A key aspect of our approach is to use an appropriate motion model that can help existing self-supervised monocular VO (SSM-VO) algorithms to overcome issues related to the local minima within their self-supervised loss functions. The motion model is expressed with a neural network named PPnet. It is trained to coarsely predict the next pose of the camera and the uncertainty of this prediction. Our self-supervised approach combines the original loss and the motion loss, which is the weighted difference between the prediction and the generated ego-motion. Taking two existing SSM-VO systems as our baseline, we evaluate our MotionHint algorithm on the standard KITTI benchmark. Experimental results show that our MotionHint algorithm can be easily applied to existing open-sourced state-of-the-art SSM-VO systems to greatly improve the performance by reducing the resulting ATE by up to 28.73%.

This paper contributes a method to design a novel navigation planner exploiting a learning-based collision prediction network. The neural network is tasked to predict the collision cost of each action sequence in a predefined motion primitives library in the robot's velocity-steering angle space, given only the current depth image and the estimated linear and angular velocities of the robot. Furthermore, we account for the uncertainty of the robot's partial state by utilizing the Unscented Transform and the uncertainty of the neural network model by using Monte Carlo dropout. The uncertainty-aware collision cost is then combined with the goal direction given by a global planner in order to determine the best action sequence to execute in a receding horizon manner. To demonstrate the method, we develop a resilient small flying robot integrating lightweight sensing and computing resources. A set of simulation and experimental studies, including a field deployment, in both cluttered and perceptually-challenging environments is conducted to evaluate the quality of the prediction network and the performance of the proposed planner.

In this work, a non-gaited framework for legged system locomotion is presented. The approach decouples the gait sequence optimization by considering the problem as a decision-making process. The redefined contact sequence problem is solved by utilizing a Monte Carlo Tree Search (MCTS) algorithm that exploits optimization-based simulations to evaluate the best search direction. The proposed scheme has proven to have a good trade-off between exploration and exploitation of the search space compared to the state-of-the-art Mixed-Integer Quadratic Programming (MIQP). The model predictive control (MPC) utilizes the gait generated by the MCTS to optimize the ground reaction forces and future footholds position. The simulation results, performed on a quadruped robot, showed that the proposed framework could generate known periodic gait and adapt the contact sequence to the encountered conditions, including external forces and terrain with unknown and variable properties. When tested on robots with different layouts, the system has also shown its reliability.

The performance of monocular depth estimation generally depends on the amount of parameters and computational cost. It leads to a large accuracy contrast between light-weight networks and heavy-weight networks, which limits their application in the real world. In this paper, we model the majority of accuracy contrast between them as the difference of depth distribution, which we call 'Distribution drift'. To this end, a distribution alignment network (DANet) is proposed. We firstly design a pyramid scene transformer (PST) module to capture inter-region interaction in multiple scales. By perceiving the difference of depth features between every two regions, DANet tends to predict a reasonable scene structure, which fits the shape of distribution to ground truth. Then, we propose a local-global optimization (LGO) scheme to realize the supervision of global range of scene depth. Thanks to the alignment of depth distribution shape and scene depth range, DANet sharply alleviates the distribution drift, and achieves a comparable performance with prior heavy-weight methods, but uses only 1% floating-point operations per second (FLOPs) of them. The experiments on two datasets, namely the widely used NYUDv2 dataset and the more challenging iBims-1 dataset, demonstrate the effectiveness of our method. The source code is available at https://github.com/YiLiM1/DANet.

In providing physical assistance to elderly people, ensuring cooperative behavior from the elderly persons is a critical requirement. In sit-to-stand assistance, for example, an older adult must lean forward, so that the body mass can shift towards the feet before a caregiver starts lifting the body. An experienced caregiver guides the older adult through verbal communications and physical interactions, so that the older adult may be cooperative throughout the process. This guidance is of paramount importance and is a major challenge in introducing a robotic aid to the eldercare environment. The wide-scope goal of the current work is to develop an in-telligent eldercare robot that can a) monitor the mental state of an older adult, and b) guide the older adult through an assisting procedure so that he/she can be cooperative in being assisted. The current work presents a basic modeling framework for describing a human's physical behaviors reflecting an internal mental state, and an algorithm for estimating the mental state through interactive observations. The sit-to-stand assistance problem is considered for the initial study. A simple Kalman Filter is constructed for estimating the level of cooperativeness in response to applied cues, with a thresholding scheme being used to make judgments on the cooperativeness state.

Modular robots are made up of a set of components which can be configured and reconfigured to form customized robots for a wide range of tasks. Fully utilizing the flexibility of modular robots is challenging, as it requires the identification of optimal modular designs for each given task, often with limited computation and time. Previous works in design automation achieve efficient run-times by utilizing machine learning to create a one-to-one mapping from task to design. However, the problem of robot design is often multimodal, where multiple distinct designs can be similarly or equally good for a task. Alternative design solutions may be needed in the field, for instance, if a module in the optimal design fails and no replacement is available. This paper presents a novel method based on generative adversarial networks (GANs) that learns a one-to-many mapping from task to a distribution of designs. We apply our method to construct locomoting modular robots for terrains with varying obstacle heights and infill. We compare our method against the state-of-the-art, and find that our algorithm results in better solution quality, diversity, and alternatives for when the optimal design fails.

This paper describes the development of a modular end-effector system (MEES) for autonomous robotic maintenance and repair tasks. The design consists of the following major components: Robot Side Mating Socket Module (RSMS), End-Effector Side Mating Socket Module (EEMS), the Modular Camera System (MCS), and Tool Holder/Changer unit. Multiple prototypes for each component have been manufactured, tested, and evaluated resulting in the final concept. Existing robotic tool-changer systems on the market were evaluated and features were built into the Modular End-Effector System to overcome the current limitations of those systems. A notable advantage to the MEES is that it is a robot agnostic system and simply uses an ISO standard bolt mounting pattern to physically attached to the robot of choice along with an ethernet connection. No external cables are required that could restrain the workspace of the robot manipulator. Additionally, it is compatible with customized wrist-mounted sensors and end-effectors without any modification of the actual robot circuitry. The MEES is demonstrated working with three different end-effectors and two different robots.

Deep imitation learning requires many expert demonstrations, which can be hard to obtain, especially when many tasks are involved. However, different tasks often share similarities, so learning them jointly can greatly benefit them and alleviate the need for many demonstrations. But, joint multi-task learning often suffers from negative transfer, sharing information that should be task-specific. In this work, we introduce a method to perform multi-task imitation while allowing for task-specific features. This is done by using proto-policies as modules to divide the tasks into simple sub-behaviours that can be shared. The proto-policies operate in parallel and are adaptively chosen by a selector mechanism that is jointly trained with the modules. Experiments on different sets of tasks show that our method improves upon the accuracy of single agents, task-conditioned and multi-headed multi-task agents, as well as state-of-the-art meta learning agents. We also demonstrate its ability to autonomously divide the tasks into both shared and task-specific sub-behaviours.

Stabilising an inverted pendulum on a cart is a well-known control problem. This paper proposes the mechan-ical and control design for solving the oscillation problem of a variable-length flexible beam mounted on a mobile robot. The system under consideration is the robot PovRob, used at the European Organization for Nuclear Research (CERN) for visual and remote inspection tasks of particle accelerators. The flexible beam mounted on the robot houses cameras and sensors. The innovative aspect of the approach concerns the use of actuated masses mounted at the end of the rod, which induces an impulsive moment due to their inertia and angular acceleration. The modelling of the flexible rod has been suitably simplified in a lumped-parameter system, with dynamic parameters related to the rod's flexibility. A linearisation of the dynamic model allows a linear-quadratic control to stabilise the system. Experimental results support the identification and the validation of the dynamic model, while simulation results evaluate the performances of the designed control law.

This paper studies a passive biped walker with knees and biomimetic feet and its behavior, as a function of key parameters. The model includes a continuous dynamic representation of the knee joint's interaction with a viscoelastic kneecap, as well as a complete kinematic description of feet that are designed to mimic the human rollover shape. First, the analytical model is derived and studied numerically for its passive walking capabilities. Then, the model is verified through independent simulations in a different platform. Finally, to increase the efficiency of its passive gaits and to map out its walking capabilities, the model is investigated parametrically. The methods used as well as the results obtained can offer significant assistance in the field of designing passivity-based biomimetic walking robots and prosthetic devices.

In this paper, we propose a new tractable ordinary differential equation formulation for dynamic simulation of fabric- reinforced inflatable soft robots. The method performs a lumped-parameter discretization of the continuum robot into discrete discs (inertia), spring elements, and threads (representing the inextensible fabric reinforcement). Using the repetition in the structure of the Lagrangian formulation of the dynamic equations of motion, a method is developed that outputs machine- readable analytical expressions for the equations of motion. The method does not require symbolic computation of derivatives. The recursive nature allows us to scale the model to an arbitrary number NN discs, and can represent buckling, twisting, and pleating that is commonly seen in very soft robots. The expressions generated were validated against manually-derived equations of motion for the two-disc case using both Lagrangian and Newton-Euler means. A simulation environment which parses and evaluates the analytical expressions generated at run-time was used to numerically integrate and predict the response of a four-disc example robot. Trajectories observed varied smoothly and plausibly predicted the behavior envisioned in robots like these.

Dielectric elastomer actuators (DEAs) have been widely employed to drive various soft robots, due to their quiet fast muscle-like behavior. It is significant but challenging to model and control these soft actuators, due to their viscoelastic property, irregular geometry, complex structure, etc. In this paper, we propose a data-driven sparse identification method to discover the hidden governing equations of DEAs. These equations can help us interpret the nonlinear properties of DEAs. Due to their low computational cost, we can further use these equations to explore classic model-based control methods for real-time accurate control of viscoelastic DEAs. The experiments show that the proposed method can model the viscoelastic behavior of the DEAs with reasonable accuracy. A feedforward controller is finally developed to validate the effectiveness of the proposed method. It is expected that this modeling method can pave the way for accurate control of soft actuators/robots with structural and material nonlinearities.

Complex assembly tasks remain huge challenge for robots because the traditional control methods rely on complicated contact state analysis. Reinforcement learning (RL) becomes one of the preferred embodiments to construct the control strategy of complex tasks. In this paper, the method of model-driven RL (MDRL) is employed to construct the control strategy. Then a completely innovative action dimension extension (ADE) mechanism is proposed to further accelerate the training process of RL. The simulation and experiment results demonstrate that the control strategy obtained through combining MDRL and ADE guarantees a more compliant assembly process. Besides, ADE method will enhance the data-efficiency of RL algorithms greatly (about 30%~40%) as well as increase the stable reward.

Comprehensive and correct state estimation with meaningful uncertainties is the basis of object-based perception for automated mobile platforms. According to fatality statistics, the most endangered group of vulnerable road users are single-track two-wheelers (ST2W), consisting mainly of cyclists, motorcyclists, and scooter riders. Due to counter-steering, they need more time to adjust their driving state to a new situation compared to four-wheelers that can directly steer in the desired direction without loosing balance. Therefore, the roll angle gives valuable information about possible future actions, especially for short, safety-critical prediction horizons. In this work, we present a basic, robust state estimation approach that is tailored to ST2W. Due to the lack of publicly available ST2W datasets with dynamic driving maneuvers and a highly accurate state, we recorded, labeled and published three different ST2W tracks ourselves. According to our results, the roll angle can be estimated bias-free with a standard deviation of between 4.4○to 6.4○, outperforming the chosen baseline.

Physical human-robot interaction can improve human ergonomics, task efficiency, and the flexibility of automation, but often requires application-specific methods to detect human state and determine robot response. At the same time, many potential human-robot interaction tasks involve discrete modes, such as phases of a task or multiple possible goals, where each mode has a distinct objective and human behavior. In this paper, we propose a novel method for multi-modal physical human-robot interaction that builds a Gaussian process model for human force in each mode of a collaborative task. These models are then used for Bayesian inference of the mode, and to determine robot reactions through model predictive control. This approach enables optimization of robot trajectory based on the belief of human intent, while considering robot impedance and human joint configuration, according to ergonomic- and/or task-related objectives. The proposed method reduces programming time and complexity, requiring only a low number of demonstrations (here, three per mode) and a mode-specific objective function to commission a flexible online human-robot collaboration task. We validate the method with experiments on an admittance-controlled robot, performing a collaborative assembly task with two modes where assistance is provided in full six degrees of freedom. It is shown that the developed algorithm robustly re-plans to changes in intent or robot initial position, achieving online control at 15 Hz.

Human-robot handover is a fundamental yet challenging task in human-robot interaction and collaboration. Recently, remarkable progressions have been made in human-to-robot handovers of unknown objects by using learning-based grasp generators. However, how to responsively generate smooth motions to take an object from a human is still an open question. Specifically, planning motions that take human comfort into account is not a part of the human-robot handover process in most prior works. In this paper, we propose to generate smooth motions via an efficient model-predictive control (MPC) framework that integrates perception and complex domain-specific constraints into the optimization problem. We introduce a learning-based grasp reachability model to select candidate grasps which maximize the robot's manipulability, giving it more freedom to satisfy these constraints. Finally, we integrate a neural net force/torque classifier that detects contact events from noisy data. We conducted human-to-robot handover experiments on a diverse set of objects with several users (N=4N=4) and performed a systematic evaluation of each module. The study shows that the users preferred our MPC approach over the baseline system by a large margin.