ICRA 2024 Papers

Keypoint detection and tracking in traditional image frames are often compromised by image quality issues such as motion blur and extreme lighting conditions. Event cameras offer potential solutions to these challenges by virtue of their high temporal resolution and high dynamic range. However, they have limited performance in practical applications due to their inherent noise in event data. This paper advocates fusing the complementary information from image frames and event streams to achieve more robust keypoint detection and tracking. Specifically, we propose a novel keypoint detection network that fuses the textural and structural information from image frames with the high-temporal-resolution motion information from event streams, namely FE-DeTr. The network leverages a temporal response consistency for supervision, ensuring stable and efficient keypoint detection. Moreover, we use a spatio-temporal nearest-neighbor search strategy for robust keypoint tracking. Extensive experiments are conducted on a new dataset featuring both image frames and event data captured under extreme conditions. The experimental results confirm the superior performance of our method over both existing frame-based and event-based methods. Our code, pre-trained models, and dataset are available at https://github.com/yuyangpoi/FE-DeTr.

3D coverage path planning for UAVs is a crucial problem in diverse practical applications. However, existing methods have shown unsatisfactory system simplicity, computation efficiency, and path quality in large and complex scenes. To address these challenges, we propose FC-Planner, a skeleton-guided planning framework that can achieve fast aerial coverage of complex 3D scenes without pre-processing. We decompose the scene into several simple subspaces by a skeleton-based space decomposition (SSD). Additionally, the skeleton guides us to effortlessly determine free space. We utilize the skeleton to efficiently generate a minimal set of specialized and informative viewpoints for complete coverage. Based on SSD, a hierarchical planner effectively divides the large planning problem into independent sub-problems, enabling parallel planning for each subspace. The carefully designed global and local planning strategies are then incorporated to guarantee both high quality and efficiency in path generation. We conduct extensive benchmark and real-world tests, where FC-Planner computes over 10 times faster compared to state-of-the-art methods with shorter path and more complete coverage. The source code will be made publicly available to benefit the community3. Project page: https://hkust-aerial-robotics.github.io/FC-Planner.

This paper presents a novel Fully Binary Point Cloud Transformer (FBPT) model which has the potential to be widely applied and expanded in the fields of robotics and mobile devices. By compressing the weights and activations of a 32-bit full-precision network to 1-bit binary values, the proposed binary point cloud Transformer network significantly reduces the storage footprint and computational resource requirements of neural network models for point cloud processing tasks, compared to full-precision point cloud networks. However, achieving a fully binary point cloud Transformer network, where all parts except the modules specific to the task are binary, poses challenges and bottlenecks in quantizing the activations of Q, K, V and self-attention in the attention module, as they do not adhere to simple probability distributions and can vary with input data. Furthermore, in our network, the binary attention module undergoes a degradation of the self-attention module due to the uniform distribution that occurs after the softmax operation. The primary focus of this paper is on addressing the performance degradation issue caused by the use of binary point cloud Transformer modules. We propose a novel binarization mechanism called dynamic-static hybridization. Specifically, our approach combines static binarization of the overall network model with fine granularity dynamic binarization of data-sensitive components. Furthermore, we make use of a novel hierarchical training scheme to obtain the optimal model and binarization parameters. These above improvements allow the proposed binarization method to outperform binarization methods applied to convolution neural networks when used in point cloud Transformer structures. To demonstrate the superiority of our algorithm, we conducted experiments on two different tasks: point cloud classification and place recognition. In point cloud classification, our model achieved an accuracy of 90.9%, which is only a 2.3% decrease compared... Show More

In wild environments, motion planning for mobile robots faces the challenge of local optimal path traps due to limited sensor perception range and lack of spatial awareness. Existing approaches that avoid local optimum by designing heuristic functions or high-quality global paths in wild environments are time-consuming and unstable. This work proposes F3DMP, which consists of two parts to alleviate the local optimum solution and better utilize distant terrain information. First, the entire planning framework is adapted to the three-dimensional space so that the planning result conforms to the geometric characteristics of the terrain. Second, a time allocation function based on offline reinforcement learning is proposed. This function can anticipate potential challenges or opportunities based on semantic information for the image and proactively determine a time allocation. Our planner is integrated into a complete mobile robot system and deployed to a real robot. Experiments in simulation and the real world demonstrate that our method can improve the success rate by 28% and the trajectory smoothness by 27% compared with traditional methods.

Reinforcement learning for control over continuous spaces typically uses high-entropy stochastic policies, such as Gaussian distributions, for local exploration and estimating policy gradient to optimize performance. Many robotic control problems deal with complex unstable dynamics, where applying actions that are off the feasible control manifolds can quickly lead to undesirable divergence. In such cases, most samples taken from the ambient action space generate low-value trajectories that hardly contribute to policy improvement, resulting in slow or failed learning. We propose to improve action selection in this model-free RL setting by introducing additional adaptive control steps based on Extremum-Seeking Control (ESC). On each action sampled from stochastic policies, we apply sinusoidal perturbations and query for estimated Q-values as the response signal. Based on ESC, we then dynamically improve the sampled actions to be closer to nearby optima before applying them to the environment. Our methods can be easily added in standard policy optimization to improve learning efficiency, which we demonstrate in various control learning environments.

Humans can perform parkour by traversing obstacles in a highly dynamic fashion requiring precise eye-muscle coordination and movement. Getting robots to do the same task requires overcoming similar challenges. Classically, this is done by independently engineering perception, actuation, and control systems to very low tolerances. This restricts them to tightly controlled settings such as a predetermined obstacle course in labs. In contrast, humans are able to learn parkour through practice without significantly changing their underlying biology. In this paper, we take a similar approach to developing robot parkour on a small low-cost robot with imprecise actuation and a single front-facing depth camera for perception which is low-frequency, jittery, and prone to artifacts. We show how a single neural net policy operating directly from a camera image, trained in simulation with large-scale RL, can overcome imprecise sensing and actuation to output highly precise control behavior end-to-end. We show our robot can perform a high jump on obstacles 2x its height, long jump across gaps 2x its length, do a handstand and run across tilted ramps, and generalize to novel obstacle courses with different physical properties. Parkour videos at https: //extreme-parkour.github.io/.

In this work, we are presenting an extension of the cooperative dual-task space (CDTS) in conformal geometric algebra. The CDTS was first defined using dual quaternion algebra and is a well established framework for the simplified definition of tasks using two manipulators. By integrating conformal geometric algebra, we aim to further enhance the geometric expressiveness and thus simplify the modeling of various tasks. We show this formulation by first presenting the CDTS and then its extension that is based around a cooperative pointpair. This extension keeps all the benefits of the original formulation that is based on dual quaternions, but adds more tools for geometric modeling of the dual-arm tasks. We also present how this CGACDTS can be seamlessly integrated with an optimal control framework in geometric algebra that was derived in previous work. In the experiments, we demonstrate how to model different objectives and constraints using the CGA-CDTS. Using a setup of two Franka Emika robots we then show the effectiveness of our approach using model predictive control in real world experiments.

A recent advance in vector field path following is the introduction of the Parametric Guiding Vector Field method. It allows for singularity-free vector fields with strong convergence guarantees, usable even for self-intersecting paths. However, the method requires significant gain tuning for practical use. In particular, for unbounded paths, the gains will inevitably become ill-suited for efficient path following. We propose a method to overcome this issue by introducing a dynamic step adaptation strategy, which provides additional normalization properties to the field. This allows the following of unbounded curves and reduces the number of gains to tune. The proposed improvements are verified in simulations using the PaparazziUAV software.

Recent advancements in age-related macular degeneration treatments necessitate precision delivery into the subretinal space, emphasizing minimally invasive procedures targeting the retinal pigment epithelium (RPE)-Bruch’s membrane complex without causing trauma. Even for skilled surgeons, the inherent hand tremors during manual surgery can jeopardize the safety of these critical interventions.This has fostered the evolution of robotic systems designed to prevent such tremors. These robots are enhanced by FBG sensors, which sense the small force interactions between the surgical instruments and retinal tissue. To enable the community to design algorithms taking advantage of such force feedback data, this paper focuses on the need to provide a specialized dataset, integrating optical coherence tomography (OCT) imaging together with the aforementioned force data.We introduce a unique dataset, integrating force sensing data synchronized with OCT B-scan images, derived from a sophisticated setup involving robotic assistance and OCT integrated microscopes. Furthermore, we present a neural network model for image-based force estimation to demonstrate the dataset’s applicability.

The inexorable progress of technology brought forth an era where robots increasingly integrate into human life which necessitates the understanding of human-robot interactions (HRI). This study unravels the details of HRI within interactive storytelling contexts. Through a between-subject experiment with 28 participants, we assessed response latency and utterance lengths to interactive story narrations delivered by either a human or a robot. Findings indicated that participants displayed longer response latency interacting with the robot narrator while articulating shorter utterances compared to the human condition where participants displayed longer utterances and shorter response latency. These observations suggest significant differences in cognitive and communicative strategies in human-human versus human-robot interactions. The results underscore the challenges and potential of designing social robots that are time-sensitive in interacting with humans. Future explorations should focus on the cognitive and emotional drivers behind these interactions.

Mobile physical human-robot collaboration (pHRC) using collaborative robots (cobots) and mobile robots has attracted much research attention. Many researchers have focused on improving the control performance to comply with human intentions. However, a problem that generally exists with mobile pHRC but often gets neglected is the impact of non-rigid components e.g. deformable tyres, suspension systems and uneven terrain on human interaction experience and task performance. To fullfil this current research gap, we carried out an investigation on the above-mentioned problem by altering a cobot’s base rigidity level (also referred to as base compliance level or BCL) during pHRC experiments. We explored how the task performance is affected by base compliance as well as human operator’s experience and cobot control parameters. Measurements include the human operator’s physical effort, task velocity, and task error. From the experimental results, it is discovered that base compliance has a significant impact on task accuracy as it can easily excite the system if an inadequate control strategy is deployed. Furthermore, through ANOVA, it is discovered that the influence of base compliance can be minimized and system excitation can be avoided by sufficient human operator training and the appropriate selection of cobot’s control parameters.

In this paper, we analyze the possibilities offered by Deep Learning State-of-the-Art architectures such as Transformers and Visual Transformers in generating a prediction of the human’s force in a Human-Robot collaborative object transportation task at a middle distance. We outperform our previous predictor by achieving a success rate of 93.8% in testset and 90.9% in real experiments with 21 volunteers predicting in both cases the force that the human will exert during the next 1 s. A modification in the architecture allows us to obtain a second output from the model with a velocity prediction, which allows us to improve the capabilities of our predictor if it is used to estimate the trajectory that the human-robot pair will follow. An ablation test is also performed to verify the relative contribution to performance of each input.

In the recent progress in embodied navigation and sim-to-robot transfer, modular policies have emerged as a de facto framework. However, there is more to compositionality beyond the decomposition of the learning load into modular components. In this work, we investigate a principled way to syntactically combine these components. Particularly, we propose Exploitation-Guided Exploration (XgX) where separate modules for exploration and exploitation come together in a novel and intuitive manner. We configure the exploitation module to take over in the deterministic final steps of navigation i.e. when the goal becomes visible. Crucially, an exploitation module teacher-forces the exploration module and continues driving an overridden policy optimization. XgX, with effective decomposition and novel guidance, improves the state-of-the-art performance on the challenging object navigation task from 70% to 73%. Along with better accuracy, through targeted analysis, we show that XgX is also more efficient at goal-conditioned exploration. Finally, we show sim-to-real transfer to robot hardware and XgX performs over two-fold better than the best baseline from simulation benchmarking. Project page: xgxvisnav.github.io

We propose the expert composer policy, a framework to reliably expand the skill repertoire of quadruped agents. The composer policy links pair of experts via transitions to a sampled target state, allowing experts to be composed sequentially. Each expert specializes in a single skill, such as a locomotion gait or a jumping motion. Instead of a hierarchical or mixture-of-experts architecture, we train a single composer policy in an independent process that is not conditioned on the other expert policies. By reusing the same composer policy, our approach enables adding new experts without affecting existing ones, enabling incremental repertoire expansion and preserving original motion quality. We measured the transition success rate of 72 transition pairs and achieved an average success rate of 99.99%, which is over 10% higher than the baseline random approach, and outperforms other state-of-the-art methods. Using domain randomization during training we ensure a successful transfer to the real world, where we achieve an average transition success rate of 97.22% (N=360) in our experiments.

Recent trends in robotic actuation have highlighted the need for low cost, high performance, and efficient gearing. We present an experimental study comparing pinwheel and non-pinwheel designs of cycloidal gearing. The open source designs are 3D-printable, combined with off-the-shelf components, achieving a high performance-to-cost ratio. Extensive experimental data is presented, that compares two prototypes on run-in behaviour and a number of quantitative metrics including transmission error, play, friction, and stiffness. Furthermore, we assess overall actuator performance through position control experiments, and a 10-hour endurance test. The results show strong performance characteristics, and crucially, suggest that non-pinwheel designs of cycloidal gearing can be a lower complexity and cost alternative to classical pinwheel designs, while offering similar performance.

Continual reinforcement learning, which aims to help robots acquire skills without catastrophic forgetting, obviating the need to re-learn all tasks from scratch. In order to enable lifelong acquisition of skills in robots, replay-based continual reinforcement learning has emerged as a promising research direction. These techniques replay data from previous tasks to mitigate forgetting when learning new skills. However, existing replay-based methods store poor representative experience, and the experience utilization of old tasks is inefficient. To address these issues, we propose an experience consistency distillation method for robot continual reinforcement learning to improve the data efficiency of the experience. Specifically, the experience of old tasks are distilled to obtain Markov Decision Process (MDP) data with high compression ratio and information content. To ensure consistent data distributions before and after distillation, we further utilize a Fréchet Inception Distance (FID) loss as a regularization constraint. In order to improve experience utilization efficiency, the policy is then trained using both the distilled data and current task data, with policy distillation performed based on uncertainty metrics. Our method is validated in the continual reinforcement learning simulation platform and real scene with a UR5e robot arm. Experimental results indicate that our method achieves higher success and lower buffer size requirement compared to other methods.

Sit-to-Stand (StS) is a fundamental daily activity that can be challenging for stroke survivors due to strength, motor control, and proprioception deficits in their lower limbs. Existing therapies involve repetitive StS exercises, but these can be physically demanding for therapists while assistive devices may limit patient participation and hinder motor learning. To address these challenges, this work proposes the use of two lower-limb exoskeletons to mediate physical interaction between therapists and patients during a StS rehabilitative task. This approach offers several advantages, including improved therapist-patient interaction, safety enforcement, and performance quantification. The whole body control of the two exoskeletons transmits online feedback between the two users, but at the same time assists in movement and ensures balance, and thus helping subjects with greater difficulty. In this study we present the architecture of the framework, presenting and discussing some technical choices made in the design.

Balance loss is a significant challenge in lower-limb exoskeleton applications, as it can lead to potential falls, thereby impacting user safety and confidence. We introduce a control framework for omnidirectional recovery step planning by online optimization of step duration and position in response to external forces. We map the step duration and position to a human-like foot trajectory, which is then translated into joint trajectories using inverse kinematics. These trajectories are executed via an impedance controller, promoting cooperation between the exoskeleton and the user. Moreover, our framework is based on the concept of the divergent component of motion, also known as the Extrapolated Center of Mass, which has been established as a consistent dynamic for describing human movement. This real-time online optimization framework enhances the adaptability of exoskeleton users under unforeseen forces thereby improving the overall user stability and safety. To validate the effectiveness of our approach, simulations, and experiments were conducted. Our push recovery experiments employing the exoskeleton in zero-torque mode (without assistance) exhibit an alignment with the exoskeleton’s recovery assistance mode, that shows the consistency of the control framework with human intention. To the best of our knowledge, this is the first cooperative push recovery framework for the lower-limb human exoskeleton that relies on the simultaneous adaptation of intra-stride parameters in both frontal and sagittal directions. The proposed control scheme has been validated with human subject experiments.

This paper proposes a novel, more computationally efficient method for optimizing robot excitation trajectories for dynamic parameter identification, emphasizing self-collision avoidance. This addresses the system identification challenges for getting high-quality training data associated with co-manipulated robotic arms that can be equipped with a variety of tools, a common scenario in industrial but also clinical and research contexts. Utilizing the Unified Robotics Description Format (URDF) to implement a symbolic Python implementation of the Recursive Newton-Euler Algorithm (RNEA), the approach aids in dynamically estimating parameters such as inertia using regression analyses on data from real robots. The excitation trajectory was evaluated and achieved on par criteria when compared to state-of-the-art reported results which didn’t consider self-collision and tool calibrations. Furthermore, physical Human-Robot Interaction (pHRI) admittance control experiments were conducted in a surgical context to evaluate the derived inverse dynamics model showing a 30.1% workload reduction by the NASA TLX questionnaire.