IROS 2021 Papers

Segmentation of a compound task with multiple subtasks is crucial for imitation learning. Conventional unsupervised segmentation methods focused on only reproducibility of demonstrations and did not use the property that goal-directed actions rarely occur without intention. In this paper, we propose a novel method to segment demonstrations into goal-directed actions by self-supervised learning. We use the discriminator between demonstrations and self-generated unintentional actions performed by the same body in behavioral cloning paradigm because goal-directed actions rarely occur without intention, and thus can be separated from unintentional actions. And we consider the states that cannot be reached by unintentional actions as subtask changepoints. We evaluated our method on manipulation tasks with multiple subtasks. The results indicate that our method can detect subtask changepoints more accurately than an existing unsupervised segmentation method.

We demonstrate that challenging shortest path problems can be solved via direct spline regression from a neural network, trained in an unsupervised manner (i.e. without requiring ground truth optimal paths for training). To achieve this, we derive a geometry-dependent optimal cost function whose minima guarantees collision-free solutions. Our method beats state-of-the-art supervised learning baselines for shortest path planning, with a much more scalable training pipeline, and a significant speedup in inference time.

Monocular depth inference has gained tremendous attention from researchers in recent years and remains as a promising replacement for expensive time-of-flight sensors, but issues with scale acquisition and implementation overhead still plague these systems. To this end, this work presents an unsupervised learning framework that is able to predict at-scale depth maps and egomotion, in addition to camera intrinsics, from a sequence of monocular images via a single network. Our method incorporates both spatial and temporal geometric constraints to resolve depth and pose scale factors, which are enforced within the supervisory reconstruction loss functions at training time. Only unlabeled stereo sequences are required for training the weights of our single-network architecture, which reduces overall implementation overhead as compared to previous methods. Our results demonstrate strong performance when compared to the current state-of-the-art on multiple sequences of the KITTI driving dataset and can provide faster training times with its reduced network complexity.

While an exciting diversity of new imaging devices is emerging that could dramatically improve robotic perception, the challenges of calibrating and interpreting these cameras have limited their uptake in the robotics community. In this work we generalise techniques from unsupervised learning to allow a robot to autonomously interpret new kinds of cameras. We consider emerging sparse light field (LF) cameras, which capture a subset of the 4D LF function describing the set of light rays passing through a plane. We introduce a generalised encoding of sparse LFs that allows unsupervised learning of odometry and depth. We demonstrate the proposed approach outperforming monocular, stereo and conventional techniques for dealing with 4D imagery, yielding more accurate odometry and depth maps and delivering these with metric scale. We anticipate our technique to generalise to a broad class of LF and sparse LF cameras, and to enable unsupervised recalibration for coping with shifts in camera behaviour over the lifetime of a robot. This work represents a first step toward streamlining the integration of new kinds of imaging devices in robotics applications.

In recent years, unsupervised deep learning approaches have received significant attention to estimating the depth and visual odometry (VO) from unlabelled monocular image sequences. However, their performance is limited in challenging environments due to perceptual degradation, occlusions, and rapid motions. Moreover, the existing unsupervised methods suffer from the lack of scale-consistency constraints across frames, which causes that the VO estimators fail to provide persistent trajectories over long sequences. In this study, we propose an unsupervised monocular deep VO framework that predicts a six-degrees-of-freedom pose camera motion and depth map of the scene from unlabelled RGB image sequences. We provide detailed quantitative and qualitative evaluations of the proposed framework on a) a challenging dataset collected during the DARPA Subterranean challenge 1; and b) the benchmark KITTI and Cityscapes datasets. The proposed approach significantly outperforms state-of-the-art unsupervised deep VO and depth prediction methods under perceptually degraded conditions providing better results for both pose estimation and depth recovery. Furthermore, it achieves state-of-the-art results in most of the VO and depth metrics on benchmark datasets. The presented approach is part of the solution used by the COSTAR team participating in the DARPA Subterranean Challenge.

Although instance-aware perception is a key prerequisite for many autonomous robotic applications, most of the methods only partially solve the problem by focusing solely on known object categories. However, for robots interacting in dynamic and cluttered environments, this is not realistic and severely limits the range of potential applications. Therefore, we propose a novel object instance segmentation approach that does not require any semantic or geometric information of the objects beforehand. In contrast to existing works, we do not explicitly use depth data as input, but rely on the insight that slight viewpoint changes, which for example are provided by stereo image pairs, are often sufficient to determine object boundaries and thus to segment objects. Focusing on the versatility of stereo sensors, we employ a transformer-based architecture that maps directly from the pair of input images to the object instances. This has the major advantage that instead of a noisy, and potentially incomplete depth map as an input, on which the segmentation is computed, we use the original image pair to infer the object instances and a dense depth map. In experiments in several different application domains, we show that our Instance Stereo Transformer (INSTR) algorithm outperforms current state-of-the-art methods that are based on depth maps. Training code and pretrained models are available at https://github.com/DLR-RM/instr.

Underwater scenarios are challenging for visual Simultaneous Localization and Mapping (SLAM) due to limited visibility and intermittently losing structures in image views. In this paper, we propose a visual acoustic bundle adjustment system which fuses a camera and a Doppler Velocity Log (DVL) in a graph SLAM framework for reliable underwater localization and mapping. In order to fuse the vision with the acoustic measurements, an calibration algorithm is also designed to estimate extrinsic parameters between a camera and a DVL using features detected in scenes. Experimental results in a tank and an offshore wind farm show the proposed method can achieve better robustness and localization accuracy than pure visual SLAM, especially in visually challenging scenarios, and the extrinsic calibration parameters can be accurately estimated, even when initialized with a random guess.

Hard coded robotic manipulation skills work well in known, predictable and repeatable situations. Human environments, however, are better described as dynamic, chaotic, uncertain or unstructured. Therefore, plans relying on preprogrammed trajectories are bound to fail in these settings. In order to increase robustness to uncertainty and avoid coding new skills from scratch, we can make flexible plans that execute existing autonomous primitives based on the sensed state of the environment. A key challenge of this approach is finding the sequence of primitives required to perform the desired task. This work uses a variation of a Hidden Markov Model (HMM) with an augmented particle filter to find the primitive sequence using only a reduced number of human demonstrations. The algorithm was tested on 40 demonstrations of two different manipulation tasks involving six primitives. It was seeded with a single manually labelled demonstration of each task and was able to automatically label the other 38 demonstration sequences with an average success of 81.5%. The results show improved convergence and a 9% increase in accuracy over other versions of the algorithm.

Parallel mechanisms (PMs) are commonly used for developing haptic devices due to low inertia, high rigidity and precision. However, limited workspace impedes their application for task-oriented robotic therapy which generally requires large motion ranges. To solve this problem, first, a PM- based reconfigurable asymmetric 6-DOF haptic interface was presented, and then a two-stage optimization method was proposed to make the robot implement two kinds of task-specific workspaces including gross motor tasks (GMTs) and fine motor tasks (FMTs). Optimization of this robot was conducted to pursue a compact size and high accuracy. The global conditioning index (GCI) and the occupied area of the robot were selected as the evaluation indices, where the GCI was derived using a dimensionally homogeneous Jacobian matrix. A multi-objective optimization method based on the genetic algorithm (GA) was utilized. The actual design parameters were finally defined from solutions of the Pareto front. The proposed two-stage optimization method provides a feasible solution for determining task-specific robotic workspace of the reconfigurable mechanism.

Parallel Continuum Robots (PCR) have received a lot of attention in recent years. This paper presents a new 6-degrees-of-freedom PCR derived from the conventional 3-PPSR parallel manipulator. This robot is driven by three limbs consisting of two flexible rods each and replacing the spherical and revolute joints of the original version. Each limb is mounted onto two linear axes arranged in series. To allow a direct comparison between the articulated and the continuum version, the parallel mechanism of an industrial manipulator has been replaced by an elastic structure of the same size. The simulations and the experiments show that the flexible counterpart of the manipulator is able to achieve a larger workspace, increasing the range of motion by 150% for rotations and by 157% in elevation. Moreover, the position repeatability is improved by 47% (reaching 3.4 µm) and the orientation repeatability by 57% (reaching 14.3 µrad). This can be explained by the removal of the spherical and revolute joints but also by the constant stress in the structure that acts as an anti backlash system on leadscrew actuators.

Supervision of a nominal controller, to enforce safety, is concerned with appropriately modifying the generated control inputs, if needed, in order to keep a control system within a set of safe states. An integral component in supervision is a controlled invariant set contained in the set of safe states. In this paper, we build on recent results on the computation of polytopic controlled invariant sets to present a supervision framework that computes the corrected inputs analytically and, hence, suitable for real-time control. The framework is validated on the task of quadrotor obstacle avoidance by forcing the vehicle to navigate within controlled invariant sets of the obstacle-free space. The results are experimentally demonstrated on a Crazyflie 2.0 quadrotor.

Compared with torque-control techniques, a position-controlled quadruped robot is lower cost, easier to build, and more direct to drive. However, the stiff actuation of position-controlled actuators makes it difficult for the quadruped to achieve dynamically stable locomotion. This paper presents an implementation of joint velocity programming technique to regulate the body’s moving speed and orientation for a position-controlled quadruped robot that performs trotting or pacing locomotion. The robot model is mapped to a new coordinate space in order to decouple the control of its body. In one plane of the new coordinate space, the robot is simplified to an inverted pendulum model to generate attitude and velocity tracking actions. In the other planes, body regulating problems are formulated in velocity forms and solved by designing support leg motions. The controllers in these planes are integrated to produce joint velocities that enable robust trotting and pacing locomotion at a variety of speeds and directions, despite lacking force control or feedback techniques. Physical test results as well as simulating results demonstrate control of the quadruped robot SmarQ to perform omni-directional locomotion, impact recovery, and adaptability to uneven terrains.

This paper presents a framework for trajectory planning that explicitly considers robotic traversability based on a quasi-dynamic vehicle model of a mobile robot in loose soil. The quasi-dynamic model estimates the slip effect due to wheel-terrain interaction forces regardless of solving complicated multibody dynamics. Therefore, our proposed model is computationally efficient for quantifying how the robot safely traverses each trajectory segment generated by a planning algorithm. The trajectory planning in our framework exploits a sampling-based incremental search algorithm, i.e., Closed-Loop Rapidly-Exploring Random Trees (CL-RRT). In the tree extension process of the CL-RRT, the traversability assessment based on the quasi-dynamic vehicle model excludes the trajectory segment associated with a hazardous wheel slip ratio. As a result, a trajectory generated from the proposed framework is safely traversable for the robot even in high slip terrain. Simulation results show that the proposed vehicle model can run 57K times faster than the dynamic model and predict the robot motion 3 times more accurately than the kinematic model. Multiple trials of the trajectory planning simulation show that our proposed framework incorporated with the quasi-dynamic model reduces a wheel slip ratio by about 40 % as compared with the kinematic model.

It is highly desirable for robots that work alongside humans to be able to understand instructions in natural language. Existing language conditioned imitation learning models directly predict the actuator commands from the image observation and the instruction text. Rather than directly predicting actuator commands, we propose translating the natural language instruction to a Python function which queries the scene by accessing the output of the object detector and controls the robot to perform the specified task. This enables the use of non-differentiable modules such as a constraint solver when computing commands to the robot. Moreover, the labels in this setup are significantly more informative computer programs that capture the intent of the expert rather than teleoperated demonstrations. We show that the proposed method performs better than training a neural network to directly predict the robot actions.

Deep imitation learning is promising for solving dexterous manipulation tasks because it does not require an environment model and pre-programmed robot behavior. However, its application to dual-arm manipulation tasks remains challenging. In a dual-arm manipulation setup, the increased number of state dimensions caused by the additional robot manipulators causes distractions and results in poor performance of the neural networks. We address this issue using a self-attention mechanism that computes dependencies between elements in a sequential input and focuses on important elements. A Transformer, a variant of self-attention architecture, is applied to deep imitation learning to solve dual-arm manipulation tasks in the real world. The proposed method has been tested on dual-arm manipulation tasks using a real robot. The experimental results demonstrated that the Transformer-based deep imitation learning architecture can attend to the important features among the sensory inputs, therefore reducing distractions and improving manipulation performance when compared with the baseline architecture without the self-attention mechanisms.

Grasping is one of the most fundamental problems in robotic manipulation. In recent years, with the development of data-driven methods, reinforcement learning has been used in solving robotic grasping problems. However, grasping is a long-horizon and sparse reward task, whose natural reward only appears when the task is successfully achieved. Therefore, it brings great challenges to the deployment of reinforcement learning methods. To tackle this difficulty, we propose a new method called Trajectory-based Split Hindsight Reverse Curriculum Learning. This method of reverse learning from the goal can greatly improve the learning efficiency and the final performance of the tasks. Specifically, based on referred trajectories, the agent starts to learn in a small state space near the goal and then gradually in larger state spaces until covering the entire state space. Through split hindsight experience replay, the sampled trajectory is divided into segments that match the current subspace's size; then, they are modified to successful trajectories to enable more efficient learning. In both simulation and real-world experiments, our method surpasses the existing methods and achieves the goal-oriented grasping tasks with higher success rates and better data efficiencies. The detailed experimental results can be viewed at https://youtu.be/7uNRzmRZhDk.

We present a reward-predictive, model-based learning method featuring trajectory-constrained visual attention for use in mapless, local visual navigation tasks. Our method learns to place visual attention at locations in latent image space which follow trajectories caused by vehicle control actions to later enhance predictive accuracy during planning. Our attention model is jointly optimized by the task-specific loss and additional trajectory-constraint loss, allowing adaptability yet encouraging a regularized structure for improved generalization and reliability. Importantly, visual attention is applied in latent feature map space instead of raw image space to promote efficient planning. We validated our model in visual navigation tasks of planning low turbulence, collision-free trajectories in off-road settings and hill climbing with locking differentials in the presence of slippery terrain. Experiments involved randomized procedural generated simulation and real-world environments. We found our method improved generalization and learning efficiency when compared to no-attention and self-attention alternatives.

Efficient trajectory optimization is essential for avoiding collisions in unstructured environments, but it remains challenging to have both speed and quality in the solutions. One reason is that second-order optimality requires calculating Hessian matrices that can grow with O(N2) with the number of waypoints. Decreasing the waypoints can quadratically decrease computation time. Unfortunately, fewer waypoints result in lower quality trajectories that may not avoid the collision. To have both, dense waypoints and reduced computation time, we took inspiration from recent studies on consensus optimization and propose a distributed formulation of collocated trajectory optimization. It breaks a long trajectory into several segments, where each segment becomes a subproblem of a few waypoints. These subproblems are solved classically, but in parallel, and the solutions are fused into a single trajectory with a consensus constraint that enforces continuity of the segments through a consensus update. With this scheme, the quadratic complexity is distributed to each segment and enables solving for higher-quality trajectories with denser waypoints. Furthermore, the proposed formulation is amenable to using any existing trajectory optimizer for solving the subproblems. We compare the performance of our implementation of trajectory splitting against leading motion planning algorithms and demonstrate the improved computational efficiency of our method.

Power-over-tether aircraft is an effective tool for persistent spatiotemporal monitoring of environmental phenomena. This paper presents the design and evaluation of flight trajectories for the tethered aircraft unmanned system (TAUS) sensing a dynamic temperature field. TAUS is a novel power-over-tether-based unmanned aerial system (UAS) configured for long-term, high throughput atmospheric monitoring. It is unique in that it provides position control while measuring atmospheric properties on-board the aircraft and with sensors along the tether. We validated the robotic system by conducting outdoor experiments to characterize the sensor performance against a meteorological tower. We found minimal sensing error at the corresponding altitude relative to the ground truth installation. We then used the experimental data to simulate four trajectories (Lawn-mower, Spiral, Star, and Flower) on power-tethered and untethered system models to evaluate performance factors related to trajectory selection. The analysis of the simulated data indicated that the power-tethered Star trajectory performed well concerning key performance factors when measuring changing atmospheric fields.