ICRA 2021 Papers

Navigation on challenging terrain topographies requires the understanding of robots’ locomotion capabilities to produce optimal solutions. We present an integrated framework for real-time autonomous navigation of mobile robots based on elevation maps. The framework performs rapid global path planning and optimization that is aware of the locomotion capabilities of the robot. A GPU-aided, sampling-based path planner combined with a gradient-based path optimizer provides optimal paths by using a neural network-based locomotion cost predictor which is trained in simulation. We show that our approach is capable of planning and optimizing paths three orders of magnitude faster than RRT* on GPU-enabled hardware, enabling real-time deployment on mobile platforms. We successfully evaluate the framework on the ANYmal C quadrupedal robot in both simulations and real-world environments for path planning tasks on multiple complex terrains.

This paper proposes a real-time obstacle avoidance control scheme for a 6-DOF manipulator with a tool in the end effector. The system consists of environment monitoring, robot-tool segmentation and collision-free motion planning of the manipulator. A Kinect V2 RGB-D camera is used to track obstacles including human and objects in the working environment. The K-D tree algorithm is then adopted to cluster point clouds of the tool and the obstacles. For robot-tool segmentation, we propose a method to model the tool in the end effector and predict its pose in order to solve the camera occlusion problem. For collision-free motion planning, a novel potential field algorithm is proposed to take into consideration of the pose of the tool. A virtual torque approach is proposed and added to the potential field in order to generate a smoother and shorter avoidance motion. The experimental results on a TM5-700 cobot show that the manipulator with a tool in the end effector effectively avoided an obstacle in real time and completed the assigned task. It is shown that the path length with the proposed virtual torque is shortened by 80.43% compared with the case without using the virtual torque.

Object instance detection is a highly relevant task to several robotic applications such as automated order picking, or household and hospital assistance robots. In these applications, a holistic scene labeling is often not required whereas it is sufficient to find a certain object type of interest, e.g. for picking it up. At the same time, large and continuously changing object sets are characteristic in such applications, requiring efficient model update capabilities from the object detector. Today’s monolithic multi-class detectors do not fulfill this criterion for fast and flexible model updates.This paper introduces InstanceNet, an ensemble of efficient single-class instance detectors capable of fast and incremental adaptation to new object sets. Due to a dynamic sampling-based training strategy, accurate detection models for new objects can be obtained within less than 40 minutes on a consumer GPU while only a small percentage of the existing detection models needs to be updated in a very efficient manner. The new detector has been thoroughly evaluated on the basis of a novel dataset of 100 grocery store objects.

An important capability of humans when performing dexterous precision gripping tasks is our ability to feel both the weight and slipperiness of an object in real-time, and adjust our grip force accordingly. In this paper, we present for the first time a fully-instrumented version of our PapillArray tactile sensor concept, which can sense grip force, object weight, and incipient slip and friction, all in real-time. We demonstrate the real-time estimation of friction and measurement of 3D force from PapillArray sensors mounted on each finger of a two-finger gripper, combined with a closed-loop grip-force control algorithm that dynamically applies a near-optimal grip force to avoid dropping objects of varying weight and friction. A vertical lifting task was performed using an object with varying weight and friction, and with some common household items. After intentionally adding a 20% safety margin on the target grip force, the actual grip force applied was only 9-30 % greater than that required to avoid slip. Future work will focus on incorporating real-time torque measurement into the grip force feedback control. This will significantly advance the state-of-the-art in artificial tactile sensing and bring us closer to robotic dexterity.

To solve the problem of unmanned ground vehicle leader-follower formation transportation in unstructured environment, we propose a novel target detection and tracking method based on multi-sensor fusion perception. Combined with 3D-Lidar, millimeter wave Radar and GPS/IMU, the proposed method can achieve stable target detection and continuous tracking of both static and dynamic vehicles. First, 3D-Lidar is used to detect the geometric model of the leader vehicle to complete the initialization of tracking target and it can also be assisted for target tracking. Then during the movement, the dynamic leader is mainly tracked through millimeter wave Radar as this sensor can keep tracking the same target with a constant index and effectively distinguish dynamic vehicle from other static obstacles according to relative speed estimation. In addition, by using GPS/IMU based integrated navigation, the movement trend of the leader can be derived according to the echo vehicle pose information and the relative position relationship. This is helpful to reduce the region of interest for target tracking and improve the real-time performance. In different unstructured environments, we perform the leader-follower formation transportation experiments for hundreds of kilometers. In rough terrain, the maximum tracking speed can still reach 40km/h and the maximum tracking distance can be up to 100 meters. Experiments show that the proposed method is suitable for vehicle target detection and tracking in unstructured environment. It has good robustness and high real-time performance with an average processing frame rate of 20Hz. The proposed method can be used for the formation transportation of unmanned ground vehicles to reduce labor costs.

We present a control architecture for real-time adaptation and tracking of trajectories generated using a terrain-aware trajectory optimization solver. This approach enables us to circumvent the computationally exhaustive task of online trajectory optimization, and further introduces a control solution robust to systems modeled with approximated dynamics. We train a policy using deep reinforcement learning (RL) to introduce additive deviations to a reference trajectory in order to generate a feedback-based trajectory tracking system for a quadrupedal robot. We train this policy across a multitude of simulated terrains and ensure its generality by introducing training methods that avoid overfitting and convergence towards local optima. Additionally, in order to capture terrain information, we include a latent representation of the height maps in the observation space of the RL environment as a form of exteroceptive feedback. We test the performance of our trained policy by tracking the corrected set points using a model-based whole-body controller and compare it with the tracking behavior obtained without the corrective feedback in several simulation environments, and show that introducing the corrective feedback results in increase of the success rate from 72.7% to 92.4% for tracking precomputed dynamic long horizon trajectories on flat terrain and from 47.5% to 80.3% on a complex modular uneven terrain. We also show successful transfer of our training approach to the real physical system and further present cogent arguments in support of our framework.

The ability to estimate 3D object shape from a single image is vital to robotics and manufacturing. For instance, it enables iterative trial-and-error in simulated environments. In single-view reconstruction, implicit functions have demonstrated superior results over traditional methods. However, implicit functions suffer from the heavy computation of mesh extraction. This is due to the indirect mesh extraction, where the number of evaluation points grows cubically with resolution. On the other hand, reducing the resolution results in the discretization error of marching cubes (MC). In this work, we aim to perform efficient and accurate mesh extraction from implicit functions. The idea is to directly reconstruct the decision boundary of implicit functions as a mesh by reverse tracing from the output. It eliminates the need for evaluating massive points and error-prone MC. Consequently, we propose implementing an implicit function via a composite function of a flow and Binary-coded Input Neural Network (BCINN). The boundary of BCINN is easily identifiable, and the flow is invertible. Owing to these properties, the decision boundary of the composite function can be directly and efficiently reconstructed. In our experiments, we demonstrate that the proposed method significantly improves runtime/memory efficiency, with results comparable to those of existing methods. Specifically, our method enables real-time high-quality mesh inference from a single image.

Real-time lower limbs motion or gait measurement is an important part in human-robotic interaction for the control of robotic walkers and rehabilitation devices. Laser range finder or infrared sensor that is mounted on the device has been widely used in applications. Although these sensors can provide accurate horizontal motion information of lower limbs during human walking, it is still difficult to measure the angular motion of lower limbs due to their functional principles. Using inertial measurement units (IMU) can measure the angular motion of lower limbs, but it requires a large amount of IMU units for measurements of all lower limb segments. In this study, a novel method is developed for real-time monitoring lower limbs (shanks and thighs) motion in human walking using just two shank-mounted IMUs. A pose prediction model based on multiple linear regression and Kalman filter is proposed. The root-mean-square error (RMSE) of the thigh orientation and the knee joint angle estimation in sagittal plane are 6.1 ± 1.3 and 6.8 ± 1.4 degs, respectively. The RMSE of the ankle, knee, and hip position estimation are 4.2 ± 1.3, 4.2 ± 1.1 and 3.5 ± 0.9 cm, respectively.

We present a task-and-motion planning (TAMP) algorithm robust against a human operator's cooperative or adversarial interventions. Interventions often invalidate the current plan and require replanning on the fly. Replanning can be computationally expensive and often interrupts seamless task execution. We introduce a dynamically reconfigurable planning methodology with behavior tree-based control strategies toward reactive TAMP, which takes the advantage of previous plans and incremental graph search during temporal logic-based reactive synthesis. Our algorithm also shows efficient recovery functionalities that minimize the number of replanning steps. Finally, our algorithm produces a robust, efficient, and complete TAMP solution. Our experimental results show the algorithm results in superior manipulation performance in both simulated and real-world tasks.

Complex manipulation tasks, such as rearrangement planning of numerous objects, are combinatorially hard problems. Existing algorithms either do not scale well or assume a great deal of prior knowledge about the environment, and few offer any rigorous guarantees. In this paper, we propose a novel hybrid control architecture for achieving such tasks with mobile manipulators. On the discrete side, we enrich a temporal logic specification with mobile manipulation primitives such as moving to a point, and grasping or moving an object. Such specifications are translated to an automaton representation, which orchestrates the physical grounding of the task to mobility or manipulation controllers. The grounding from the discrete to the continuous reactive controller is online and can respond to the discovery of unknown obstacles or decide to push out of the way movable objects that prohibit task accomplishment. Despite the problem complexity, we prove that, under specific conditions, our architecture enjoys provable completeness on the discrete side, provable termination on the continuous side, and avoids all obstacles in the environment. Simulations illustrate the efficiency of our architecture that can handle tasks of increased complexity while also responding to unknown obstacles or unanticipated adverse configurations.

Human-robot object handovers have been an actively studied area of robotics over the past decade; however, very few techniques and systems have addressed the challenge of handing over diverse objects with arbitrary appearance, size, shape, and deformability. In this paper, we present a vision-based system that enables reactive human-to-robot handovers of unknown objects. Our approach combines closed-loop motion planning with real-time, temporally consistent grasp generation to ensure reactivity and motion smoothness. Our system is robust to different object positions and orientations, and can grasp both rigid and non-rigid objects. We demonstrate the generalizability, usability, and robustness of our approach on a novel benchmark set of 26 diverse household objects, a user study with six participants handing over a subset of 15 objects, and a systematic evaluation examining different ways of handing objects.

This paper addresses the problem of real-time planning in constrained dual-arm manipulation scenarios. Our proposed coupling leverages manipulability information of the cooperative bimanual task-space to a vector-field based planner by means of a repulsive circulatory field, while geometric primitives in Spin(3)⋉ℝ3 are explored for flexible task relaxation. Furthermore, the circular field informs the cooperative framework about the safety boundaries which are in turn used to further relax motion constraints within a collision-free ball in Cartesian space. This builds a funnel along the trajectory which can be directly tracked through the proposed switching of task-primitive-priorities. The switching strategy follows an approach that ensures robustness to chattering and continuity in the joint-space. Experiments verify that our framework can run within the inner control loop of Franka Emika Panda robots.

We outline a neural network-based pipeline for perception, control and planning of a 7 DoF robot for tasks that involve reaching into a dormant grapevine canopy. The proposed system consists of a 6 DoF industrial robot arm and a linear slider that can actuate on an entire grape vine. Our approach uses Convolutional Neural Networks to detect buds in dormant grape vines and a Reinforcement Learning based control strategy to reach desired cut-point locations for pruning tasks. Within this framework, three methodologies are developed and compared to reach the desired locations: the learned policy-based approach (RL), a hybrid method that uses the learned policy and an inverse kinematics solver (RL+IK), and lastly a classical approach commonly used in robotics. We first tested and validated the suitability of the proposed learning methodology in a simulated environment that resembled laboratory conditions. A reaching accuracy of up to 61.90% and 85.71% for the RL and RL+IK approaches respectively was obtained for a vine that the agent observed while learning. When testing in a new vine, the accuracy was up to 66.66% and 76.19% for RL and RL+IK, respectively. The same methods were then deployed on a real system in an end to end procedure: autonomously scan the vine using a vision system, create its model and finally use the learned policy to reach cutting points. The reaching accuracy obtained in these tests was 73.08%.

We present a method for computing exact reachable sets for deep neural networks with rectified linear unit (ReLU) activation. Our method is well-suited for use in rigorous safety analysis of robotic perception and control systems with deep neural network components. Our algorithm can compute both forward and backward reachable sets for a ReLU network iterated over multiple time steps, as would be found in a perception-action loop in a robotic system. Our algorithm is unique in that it builds the reachable sets by incrementally enumerating polyhedral cells in the input space, rather than iterating layer-by-layer through the network as in other methods. If an unsafe cell is found, our algorithm can return this result without completing the full reachability computation, thus giving an anytime property that accelerates safety verification. In addition, our method requires less memory during execution compared to existing methods where memory can be a limiting factor. We demonstrate our algorithm on safety verification of the ACAS Xu aircraft advisory system. We find unsafe actions many times faster than the fastest existing method and certify no unsafe actions exist in about twice the time of the existing method. We also compute forward and backward reachable sets for a learned model of pendulum dynamics over a 50 time step horizon in 87s on a laptop computer. Algorithm source code: https://github.com/StanfordMSL/Neural-Network-Reach.

This paper studies push recovery for humanoid robots based on a variable-height inverted pendulum (VHIP) model. We first develop an approach for treating zero-step capturability of the VHIP with a novel methodology based on Hamilton-Jacobi (HJ) reachability analysis. Such an approach uses the sub-zero level set of a value function to encode capturability of the VHIP, where the value function is obtained by numerically solving a HJ variational inequality offline. Based on this analysis, a simple and effective method for adjusting foothold locations is then devised for cases where the VHIP state is not zero-step capturable. In addition, the HJ reachability analysis naturally induces an optimal control law that allows for rapid planning with the VHIP during push recovery online. To enable use of the strategy with a position-controlled humanoid robot, an associated differential inverse kinematics based tracking controller is employed. The effectiveness of the overall framework is demonstrated with the UBTECH Walker robot in the MuJoCo simulator. Simulation validations show a significant improvement in push robustness as compared to the methods based on the classical linear inverted pendulum model.

Motivated by earlier work and the developer of a new algorithm, the FollowerStopper, this article uses reachability analysis to verify the safety of the FollowerStopper algorithm, which is a controller designed for dampening stop-and-go traffic waves. With more than 1100 miles of driving data collected by our physical platform, we validate our analysis results by comparing it to human driving behaviors. The FollowerStopper controller has been demonstrated to dampen stop-and-go traffic waves at low speed, but previous analysis on its relative safety has been limited to upper and lower bounds of acceleration. To expand upon previous analysis, reachability analysis is used to investigate the safety at the speeds it was originally tested and also at higher speeds. Two formulations of safety analysis with different criteria are shown: distance-based and time headway-based. The FollowerStopper is considered safe with distance-based criterion. However, simulation results demonstrate that the FollowerStopper is not representative of human drivers - it follows too closely behind vehicles, specifically at a distance human would deem as unsafe. On the other hand, under the time headway-based safety analysis, the FollowerStopper is not considered safe anymore. A modified FollowerStopper is proposed to satisfy time-based safety criterion. Simulation results of the proposed FollowerStopper shows that its response represents human driver behavior better.

Many Reinforcement Learning (RL) approaches use joint control signals (positions, velocities, torques) as action space for continuous control tasks. We propose to lift the action space to a higher level in the form of subgoals for a motion generator (a combination of motion planner and trajectory executor). We argue that, by lifting the action space and by leveraging sampling-based motion planners, we can efficiently use RL to solve complex, long-horizon tasks that could not be solved with existing RL methods in the original action space. We propose ReLMoGen – a framework that combines a learned policy to predict subgoals and a motion generator to plan and execute the motion needed to reach these subgoals. To validate our method, we apply ReLMoGen to two types of tasks: 1) Interactive Navigation tasks, navigation problems where interactions with the environment are required to reach the destination, and 2) Mobile Manipulation tasks, manipulation tasks that require moving the robot base. These problems are challenging because they are usually long-horizon, hard to explore during training, and comprise alternating phases of navigation and interaction. Our method is benchmarked on a diverse set of seven robotics tasks in photo-realistic simulation environments. In all settings, ReLMoGen outperforms state-of-the-art RL and Hierarchical RL baselines. ReLMoGen also shows outstanding transferability between different motion generators at test time, indicating a great potential to transfer to real robots. For more information, please visit project website: http://svl.stanford.edu/projects/relmogen.

Recent advances in machine learning have triggered an enormous interest in using learning-based approaches for robot control and object manipulation. While the majority of existing algorithms are evaluated under the assumption that the involved bodies are rigid, a large number of practical applications contain deformable objects. In this work we focus on Deformable Linear Objects (DLOs) which can be used to model cables, tubes or wires. They are present in many applications such as manufacturing, agriculture and medicine. New methods in robotic manipulation research are often demonstrated in custom environments impeding reproducibility and comparisons of algorithms. We introduce ReForm, a simulation sandbox and a tool for benchmarking manipulation of DLOs. We offer six distinct environments representing important characteristics of deformable objects such as elasticity, plasticity or self-collisions and occlusions. A modular framework is used, enabling design parameters such as the end-effector degrees of freedom, reward function and type of observation. ReForm is a novel robot learning sandbox with which we intend to facilitate testing and reproducibility in manipulation research for DLOs.

Bilateral teleoperation under rate mode is known to be a difficult problem in terms of stability, especially when the slave manipulator interacts with a time-varying environment. This paper presents an energy based variable admittance control approach, whose principle combines the monitoring and the regulation of the energy exchanges with a passivity tank. It allows stable interactions with force feedback for any desired inertia, damping and stiffness parameters. Experiments are conducted to assess the efficiency of the proposed approach using an experimental setup with a variable stiffness environment. The obtained results illustrate the ability of the proposed strategy to stabilize a system otherwise unstable, with little effect on the transparency of the teleoperation system.