ICRA 2021 Papers

This paper introduces a real-time dense planar LiDAR SLAM system, named π-LSAM, for the indoor environment. The widely used LiDAR odometry and mapping (LOAM) framework [1] does not include bundle adjustment (BA) and generates a low fidelity tracking pose. This paper seeks to overcome these drawbacks for the indoor environment. Specifically, we use the plane as the landmark, and introduce plane adjustment (PA) as our back-end to jointly optimize planes and keyframe poses. We present the π-factor to significantly reduce the computational complexity of PA. In addition, we introduce an efficient loop detection algorithm based on the RANSAC framework using planes. In the front-end, our algorithm performs global registration in real time. To achieve this performance, we maintain the local-to-global point-to-plane correspondences scan by scan, so that we only need a small local KD-tree to establish the data association between a LiDAR scan and the global planes, rather than a large global KD-tree used in previous works. With this local-to-global data association, our algorithm directly identifies planes in a LiDAR scan, and yields an accurate and globally consistent pose. Experimental results show that our algorithm significantly outperforms the state-of-the-art LOAM variant, LeGO-LOAM [2], and our algorithm achieves real time.

While traditional approaches to manipulation planning assume known object templates, recent approaches to "category-level manipulation" aim to manipulate a category of objects with potentially unknown instances and large intra-category shape variation. In this paper we explore an object representation to enable precise category-level manipulation, capturing a notion of the object configuration and extent, while being generalizable to novel instances. Building on our previous work, kPAM 1, we combine semantic keypoints with dense geometry (a point cloud or mesh) as the interface between the perception module and motion planner. Leveraging advances in learning-based keypoint detection and shape completion, both dense geometry and keypoints can be perceived from raw sensor input. Using the proposed hybrid object representation, we formulate the manipulation task as a motion planning problem which encodes both the object target configuration and physical feasibility for a category of objects. In this way, many existing manipulation planners can be generalized to categories of objects, and the resulting perception-to-action manipulation pipeline is robust to large intra-category shape variation. Extensive hardware experiments demonstrate our pipeline can produce robot trajectories that accomplish tasks with never-before-seen objects. The video demo is available on this link: https://sites.google.com/view/generalizable-manipulation.

In recent years, there have been significant advances in building end-to-end Machine Learning (ML) systems that learn at scale. But most of these systems are: (a) isolated (perception, speech, or language only); (b) trained on static datasets. On the other hand, in the field of robotics, large-scale learning has always been difficult. Supervision is hard to gather and real world physical interactions are expensive.In this work we introduce and open-source droidlet, a modular, heterogeneous agent architecture and platform. It allows us to exploit both large-scale static datasets in perception and language and sophisticated heuristics often used in robotics; and provides tools for interactive annotation. Furthermore, it brings together perception, language and action onto one platform, providing a path towards agents that learn from the richness of real world interactions.

It is a long-standing challenge to enable an intelligent agent to learn in one environment and generalize to an unseen environment without further data collection and finetuning. In this paper, we consider a zero shot generalization problem setup that complies with biological intelligent agents’ learning and generalization processes. The agent is first presented with previous experiences in the training environment, along with task description in the form of trajectory-level sparse rewards. Later when it is placed in the new testing environment, it is asked to perform the task without any interaction with the testing environment. We find this setting natural for biological creatures and at the same time, challenging for previous methods. Behavior cloning, state-of-art RL along with other zero-shot learning methods perform poorly on this benchmark. Given a set of experiences in the training environment, our method learns a neural function that decomposes the sparse reward into particular regions in a contingency-aware observation as a per step reward. Based on such decomposed rewards, we further learn a dynamics model and use Model Predictive Control (MPC) to obtain a policy. Since the rewards are decomposed to finer-granularity observations, they are naturally generalizable to new environments that are composed of similar basic elements. We demonstrate our method on a wide range of environments, including a classic video game – Super Mario Bros, as well as a robotic continuous control task. Please refer to the project page for more visualized results.1

This paper studies the problem of autonomous exploration under localization uncertainty for a mobile robot with 3D range sensing. We present a framework for self-learning a high-performance exploration policy in a single simulation environment, and transferring it to other environments, which may be physical or virtual. Recent work in transfer learning achieves encouraging performance by domain adaptation and domain randomization to expose an agent to scenarios that fill the inherent gaps in sim2sim and sim2real approaches. However, it is inefficient to train an agent in environments with randomized conditions to learn the important features of its current state. An agent can use domain knowledge provided by human experts to learn efficiently. We propose a novel approach that uses graph neural networks in conjunction with deep reinforcement learning, enabling decision-making over graphs containing relevant exploration information provided by human experts to predict a robot's optimal sensing action in belief space. The policy, which is trained only in a single simulation environment, offers a real-time, scalable, and transferable decision-making strategy, resulting in zero-shot transfer to other simulation environments and even real-world environments.

This paper presents an analytical methodology and experimental study to identify quantitatively the zero-potential-energy (ZP) motion due to the stiffness matrices in Cartesian impedance control of redundant manipulators. This mode of motion, analogous to the rigid-body mode in classic mechanical systems, shows up as a result of the redundancy of the robot and creates a steady-state deviation from its initial configuration after it reaches equilibrium when subject to a perturbation, because of the principle of least energy for dynamic systems. We determine such ZP motion(s) by utilizing a vibration-based closed-form solution recently developed. We identify and provide experimental validation of the existence of the ZP motions on a 7 DoF Panda robot.

Pose estimation is a basic module in many robot manipulation pipelines. Estimating the pose of objects in the environment can be useful for grasping, motion planning, or manipulation. However, current state-of-the-art methods for pose estimation either rely on large annotated training sets or simulated data. Further, the long training times for these methods prohibit quick interaction with novel objects. To address these issues, we introduce a novel method for zero-shot object pose estimation in clutter. Our approach uses a hypothesis generation and scoring framework, with a focus on learning a scoring function that generalizes to objects not used for training. We achieve zero-shot generalization by rating hypotheses as a function of unordered point differences. We evaluate our method on challenging datasets with both textured and untextured objects in cluttered scenes and demonstrate that our method significantly outperforms previous methods on this task. We also demonstrate how our system can be used by quickly scanning and building a model of a novel object, which can immediately be used by our method for pose estimation. Our work allows users to estimate the pose of novel objects without requiring any retraining. Additional information can be found on our website https://bokorn.github.io/zephyr/

We propose YolactEdge, the first competitive instance segmentation approach that runs on small edge devices at real-time speeds. Specifically, YolactEdge runs at up to 30.8 FPS on a Jetson AGX Xavier (and 172.7 FPS on an RTX 2080 Ti) with a ResNet-101 backbone on 550x550 resolution images. To achieve this, we make two improvements to the state-of-the-art image-based real-time method YOLACT [1]: (1) applying TensorRT optimization while carefully trading off speed and accuracy, and (2) a novel feature warping module to exploit temporal redundancy in videos. Experiments on the YouTube VIS and MS COCO datasets demonstrate that YolactEdge produces a 3-5x speed up over existing real-time methods while producing competitive mask and box detection accuracy. We also conduct ablation studies to dissect our design choices and modules. Code and models are available at https://github.com/haotian-liu/yolact_edge.

Object detection in 3D with stereo cameras is an important problem in computer vision, and is particularly crucial in low-cost autonomous mobile robots without LiDARs. Nowadays, most of the best-performing frameworks for stereo 3D object detection are based on dense depth reconstruction from disparity estimation, making them extremely computationally expensive. To enable real-world deployments of vision detection with binocular images, we take a step back to gain insights from 2D image-based detection frameworks and enhance them with stereo features. We incorporate knowledge and the inference structure from real-time one-stage 2D/3D object detector and introduce a light-weight stereo matching module. Our proposed framework, YOLOStereo3D, is trained on one single GPU and runs at more than ten fps. It demonstrates performance comparable to state-of-the-art stereo 3D detection frameworks without usage of LiDAR data. The code will be published in https://github.com/Owen-Liuyuxuan/visualDet3D.

Simulation is at the core of validating autonomous systems (AS), enabling the detection of faults at a lower cost and earlier in the development life cycle. However, simulation can only produce an approximation of the real world, leading to a gap between simulation and reality where undesirable system behaviors can go unnoticed. To address that gap, we present a novel approach, world-in-the-loop (WIL) simulation, which integrates sensing data from simulation and the real world to provide the AS with a mixed-reality. The approach executes multiple instances of the AS in parallel, one in the real world and at least one in simulation, performs configurable transformations, filtering, and merging operations on the body of sensed data in order to integrate it, and provides the pipelines to distribute the original sensor data and the integrated sensor data back to the executing AS. We present a study on multiple scenarios and two simulators that demonstrates how WIL reduces the simulation-reality gap and increases the chances of exposing failures before deployment.

Multicopters are able to perform high maneuverability yet their potential have not been fully achieved. In this work, we propose a full-body, optimization-based motion planning framework that takes shape and attitude of aerial robot into consideration such that the aggressiveness of drone maneuvering improves significantly in cluttered environment. Our method takes in a series of intersecting polyhedrons that describe a range of 3D free spaces and outputs a time-indexed trajectory in real-time with full-body collision-free guarantee. The drone is modeled as a tilted cuboid, yet we argue that our framework can be freely adjusted to fit multicopters of different shapes. Guaranteeing dynamic feasibility and safety conditions, our framework transforms the original constrained nonlinear programming problem to an unconstrained one in higher dimensions which is further solved by quasi-Newton methods. Benchmark has shown that our method improves the state-of-art with orders of magnitude in terms of computation time and memory usage. Simulations and onboard experiments are carried out as validation.

This paper presents the first successful experiment implementing whole-body model predictive control with state feedback on a torque-control humanoid robot. We demonstrate that our control scheme is able to do whole-body target tracking, control the balance in front of strong external perturbations and avoid collision with an external object. The key elements for this success are threefold. First, optimal control over a receding horizon is implemented with Crocoddyl, an optimal control library based on differential dynamics programming, providing state-feedback control in less than 10 ms. Second, a warm start strategy based on memory of motion has been implemented to overcome the sensitivity of the optimal control solver to initial conditions. Finally, the optimal trajectories are executed by a low-level torque controller, feedbacking on direct torque measurement at high frequency. This paper provides the details of the method, along with analytical benchmarks with the real humanoid robot Talos.A video of the experiment is available at https://peertube.laas.fr/videos/watch/cbc25927-337c-4635-a1bc-153b9aeb4135

Talking gestures are a fundamental part of body language and, therefore, are also important for social robots. Gesture generation by generative approaches is supposed to produce a more appropriate behavior than rule-based approaches. Usually, the evaluation of generated gestures is carried out by subjective visual evaluation, which could be cultural dependent and influenced by external factors. In this work we extend previous research on quantitative evaluation methods, comparing two generative methods and showing that their results correlate with subjective evaluation by a sizable group of people. The final goal is to offer a quantitative tool to help the researchers to automate the evaluation of their gesture generation systems, as a complementary measure to subjective methods.

Human-robot interactions (HRI) can be modeled as differential games with incomplete information, where each agent holds private reward parameters. Due to the open challenge in finding perfect Bayesian equilibria of such games, existing studies often decouple the belief and physical dynamics by iterating between belief update and motion planning. Importantly, the robot’s reward parameters are often assumed to be known to the humans, in order to simplify the computation. We show in this paper that under this simplification, the robot performs non-empathetic belief update about the humans’ parameters, which causes high safety risks in uncontrolled intersection scenarios. In contrast, we propose a model for empathetic belief update, where the agent updates the joint probabilities of all agents’ parameter combinations. The update uses a neural network that approximates the Nash equilibrial action-values of agents. We compare empathetic and non-empathetic belief update methods on a two-vehicle uncontrolled intersection case with short reaction time. Results show that when both agents are unknowingly aggressive (or non-aggressive), empathy is necessary for avoiding collisions when agents have false believes about each others’ parameters. This paper demonstrates the importance of acknowledging the incomplete-information nature of HRI.

We investigate what grade of sensor data is required for training an imitation-learning-based AV planner on human expert demonstration. Machine-learned planners [1] are very hungry for training data, which is usually collected using vehicles equipped with the same sensors used for autonomous operation [1]. This is costly and non-scalable. If cheaper sensors could be used for collection instead, data availability would go up, which is crucial in a field where data volume requirements are large and availability is small. We present experiments using up to 1000 hours worth of expert demonstration and find that training with 10x lower-quality data outperforms 1x AV-grade data in terms of planner performance (see Fig. 1). The important implication of this is that cheaper sensors can indeed be used. This serves to improve data access and democratize the field of imitation-based motion planning. Alongside this, we perform a sensitivity analysis of planner performance as a function of perception range, field-of-view, accuracy, and data volume, and reason about why lower-quality data still provide good planning results.

The estimation of the orientation of an observed vehicle relative to an Autonomous Vehicle (AV) from monocular camera data is an important building block in estimating its 6 DoF pose. Current Deep Learning based solutions for placing a 3D bounding box around this observed vehicle are data hungry and do not generalize well. In this paper, we demonstrate the use of monocular visual odometry for the self-supervised fine-tuning of a model for orientation estimation pre-trained on a reference domain. Specifically, while transitioning from a virtual dataset (vKITTI) to nuScenes, we recover up to 70% of the performance of a fully supervised method. We subsequently demonstrate an optimization-based monocular 3D bounding box detector built on top of the self-supervised vehicle orientation estimator without the requirement of expensive labeled data. This allows 3D vehicle detection algorithms to be self-trained from large amounts of monocular camera data from existing commercial vehicle fleets.

A generalist robot equipped with learned skills must be able to perform many tasks in many different environments. However, zero-shot generalization to new settings is not always possible. When the robot encounters a new environment or object, it may need to finetune some of its previously learned skills to accommodate this change. But crucially, previously learned behaviors and models should still be suitable to accelerate this relearning. In this paper, we aim to study how generative models of possible outcomes can allow a robot to learn visual representations of affordances, so that the robot can sample potentially possible outcomes in new situations, and then further train its policy to achieve those outcomes. In effect, prior data is used to learn what kinds of outcomes may be possible, such that when the robot encounters an unfamiliar setting, it can sample potential outcomes from its model, attempt to reach them, and thereby update both its skills and its outcome model. We show that this approach can be used to train goal-conditioned policies that operate on raw image inputs, and can rapidly learn to manipulate new objects via our proposed affordance-directed exploration scheme.

Soil strength testing and collecting soil cores from wetlands is currently a slow, manual process that runs the risk of disturbing and contaminating soil samples. This paper describes a method using an instrumented dart deployed and retrieved by a drone for performing core sample tests in soft soils. The instrumented dart can simultaneously conduct free- fall penetrometer tests. A drone-mounted mechanism enables deploying and reeling in the dart for sample return or for multiple soil strength tests. Tests examine the effect of dart tip diameter and drop height on soil retrieval, and the requisite pull force to retrieve the samples. Further tests examine the dart’s ability to measure soil strength and penetration depth. Hardware trials demonstrate that the drone can repeatedly drop and retrieve a dart, and that the soil can be discretely sampled.

This paper implements and demonstrates visual route mapping and localisation upon a Pixel Processor Array (PPA). The PPA sensor comprises of an array of Processing Elements (PEs), each of which can capture and process visual information directly. This provides significant parallel processing power allowing novel ways in which information can be processed on-sensor. Our method predicts the correct node within a topological map generated from an image sequence by measuring image similarities, spatial coherence, and exploiting the parallel nature of the PPA. Our implementation runs at +300Hz on large public datasets with +2K locations requiring 2.5W at 500 GOPS/W. We compare vs traditionally implemented methods demonstrating better F-1 performance even on simulation. As far as we are aware, we present the first on-sensor mapping and localisation system running entirely on-sensor.