ICRA 2024 Papers

Recent work has shown the promise of creating generalist, transformer-based, models for language, vision, and sequential decision-making problems. To create such models, we generally require centralized training objectives, data, and compute. It is of interest if we can more flexibly create generalist policies by merging together multiple, task-specific, individually trained policies. In this work, we take a preliminary step in this direction through merging, or averaging, subsets of Decision Transformers in parameter space trained on different MuJoCo locomotion problems, forming multi-task models without centralized training. We also demonstrate the importance of various methodological choices when merging policies, such as utilizing common pre-trained initializations, increasing model capacity, and utilizing Fisher information for weighting parameter importance. In general, we believe research in this direction could help democratize and distribute the process that forms multi-task robotics policies. Our implementation is available at https://github.com/daniellawson9999/merging-decision-transformer.

This paper presents a 6-DoF range-based Monte Carlo localization method with a GPU-accelerated Stein particle filter. To update a massive amount of particles, we propose a Gauss-Newton-based Stein variational gradient descent (SVGD) with iterative neighbor particle search. This method uses SVGD to collectively update particle states with gradient and neighborhood information, which provides efficient particle sampling. For an efficient neighbor particle search, it uses locality sensitive hashing and iteratively updates the neighbor list of each particle over time. The neighbor list is then used to propagate the posterior probabilities of particles over the neighbor particle graph. The proposed method is capable of evaluating one million particles in real-time on a single GPU and enables robust pose initialization and re-localization without an initial pose estimate. In experiments, the proposed method showed an extreme robustness to complete sensor occlusion (i.e., kidnapping), and enabled pinpoint sensor localization without any prior information.

Powerline robotics is slowly becoming key tools for electric utilities. Contrary to drones that are usually limited to inspection tasks, wheeled robots like LineRanger can perform a broader range of applications. In this paper, a suite of mechanical devices is featured, as several new asset management tasks were recently added to LineRanger’s capabilities. While previous applications focused on non-contact inspection (visual, electro-magnetic, etc.), the new tasks at hand involved reaching adjacent conductors to probe line components with micro-Ohmmeter, installing and retrieving custom build sensors for multi-day line monitoring, and assessing aging conductors surface properties, to refine their thermal model and optimize the line capacity during heat waves. All three applications were recently field validated onto LineRanger, and mechanical design insights shall be presented for each module.

A two-wheeled vehicle has a compact structure and high mobility in crowded and complex environments, which has been widely used in urban logistics. The bicycle and self-balancing vehicle are the two main modes of the two-wheeled vehicle, and their combination allows for good balance-control stability at both high and low speeds. Four control inputs by two steerable driving wheels are required to implement transformations between the two modes due to the difference in their configuration spaces. However, the control inputs are redundant for planar motions, which results in an over-constraint of the vehicle. In this work, a two-wheeled robot with an additional structural degree of freedom (DOF) is designed to balance inputs and DOFs to avoid over-constraint. A transition mode based on oblique vehicle motions is used to bridge the transformation of the bicycle and self-balancing vehicle modes. A general kinematic model is developed for the two-wheeled robot’s planar motions, and the three modes’ kinematics are special cases with particular servo constraints. Structural DOF control laws are developed and experimentally validated on a prototype robot. Smooth transformations of the multimodal motions are also validated by using the prototype.

Automatic inspection of parabolic-trough solar plants is key to preventing failures that can harm the environment and the production of green energy. In this work, we propose a novel methodology to inspect ball joints in parabolic trough collectors, which is a relevant problem that is not adequately covered in the literature. Images collected by an Unmanned Aerial Vehicle are segmented using deep learning to extract ball joint components. In order to generate rich training datasets, we develop a novel data augmentation technique by rotating joints and adding synthetic image background, and demonstrate its impact on the object detection accuracy. Then two types of faults are analyzed: fluid leaks, by means of image color filtering; and geometric shape anomalies, by measuring joint angles of the robotic arms. We propose metrics to quantify these faults and evaluate the damage of the inspected components. Our experimental results with images from operating commercial plants show that we can automatically detect leaks and anomalous angular geometry with a low failure rate compared to human labeling.

Relative navigation methods are a critical enabling technology for the next generation of autonomous spacecraft conducting close proximity operations. This is especially true for multi-agent inspection operations in which safety including intra-agent or agent-target collisions are a serious concern. Additionally, in an on-orbit servicing operation various failure modes of the target may result in unreliable a-priori knowledge or cooperation from the target. The main contribution of this work is the demonstration of a method for multi-agent, relative pose estimation that is robust to A) sensor blinding and B) dynamic uncertainty. This objective is accomplished leveraging GTSAM, an existing toolbox for the formulation of factor graphs, along with an algorithm for the efficient, real-time solution of such factor graphs, iSAM2. This estimation method is demonstrated in an example scenario with uncertain dynamics and sensor blinding due to sun position. Results revealed that the iSAM2-based method is capable of handling sensor blinding through leveraging an inter-agent range measurement, despite a dynamically uncertain environment.

Segmenting unseen objects from images is a critical perception skill that a robot needs to acquire. In robot manipulation, it can facilitate a robot to grasp and manipulate unseen objects. Mean shift clustering is a widely used method for image segmentation tasks. However, the traditional mean shift clustering algorithm is not differentiable, making it difficult to integrate it into an end-to-end neural network training framework. In this work, we propose the Mean Shift Mask Transformer (MSMFormer), a new transformer architecture that simulates the von Mises-Fisher (vMF) mean shift clustering algorithm, allowing for the joint training and inference of both the feature extractor and the clustering. Its central component is a hypersphere attention mechanism, which updates object queries on a hypersphere. To illustrate the effectiveness of our method, we apply MSMFormer to unseen object instance segmentation. Our experiments show that MSMFormer achieves competitive performance compared to state-of-the-art methods for unseen object instance segmentation 1.

Reinforcement Learning (RL) has been a powerful tool for training robots to acquire agile locomotion skills. To learn locomotion, it is commonly necessary to introduce additional reward-shaping terms, such as an energy minimization term, to guide an algorithm like Proximal Policy Optimization (PPO) to good performance. Prior works rely on hyper-parameter tuning on the weight of the reward shaping terms to obtain satisfactory task performance. To save the efforts of tuning these weights, we adopt the Extrinsic-Intrinsic Policy Optimization (EIPO) framework. The key idea of EIPO is to establish a constrained optimization framework for the primary objective of enhancing task performance and the secondary objective of minimizing energy consumption. It seeks a policy that minimizes the energy consumption objective within the optimal policy space for task performance. This guarantees that the learned policy excels in task performance while conserving energy, all without requiring manual weight adjustments for both objectives. Our experiments evaluate EIPO on various quadruped locomotion tasks, revealing that policies trained with EIPO consistently achieve higher task performance than PPO comparisons while maintaining comparable energy consumption levels. Furthermore, EIPO exhibits superior task performance in real-world evaluations compared to PPO.

People with Visual Impairments (PVI) typically recognize objects through haptic perception. Knowing objects and materials before touching is desired by the target users but under-explored in the field of human-centered robotics. To fill this gap, in this work, a wearable vision-based robotic system, MATERobot, is established for PVI to recognize materials and object categories beforehand. To address the computational constraints of mobile platforms, we propose a lightweight yet accurate model MATEViT to perform pixel-wise semantic segmentation, simultaneously recognizing both objects and materials. Our methods achieve respective 40.2% and 51.1% of mIoU on COCOStuff-10K and DMS datasets, surpassing the previous method with +5.7% and +7.0% gains. Moreover, on the field test with participants, our wearable system reaches a score of 28 in the NASA-Task Load Index, indicating low cognitive demands and ease of use. Our MATERobot demonstrates the feasibility of recognizing material property through visual cues and offers a promising step towards improving the functionality of wearable robots for PVI. The source code has been made publicly available at MATERobot.

Reinforcement learning solely from an agent’s self-generated data is often believed to be infeasible for learning on real robots, due to the amount of data needed. However, if done right, agents learning from real data can be surprisingly efficient through re-using previously collected sub-optimal data. In this paper we demonstrate how the increased understanding of off-policy learning methods and their embedding in an iterative online/offline scheme ("collect and infer") can drastically improve data-efficiency by using all the collected experience, which empowers learning from real robot experience only. Moreover, the resulting policy improves significantly over the state of the art on a recently proposed real robot manipulation benchmark. Our approach learns end-to-end, directly from pixels, and does not rely on additional human domain knowledge such as a simulator or demonstrations.

This work proposes a semantic segmentation network that produces high-quality uncertainty estimates in a single forward pass. We exploit general representations from foundation models and unlabelled datasets through a Masked Image Modeling (MIM) approach, which is robust to augmentation hyper-parameters and simpler than previous techniques. For neural networks used in safety-critical applications, bias in the training data can lead to errors; therefore it is crucial to understand a network’s limitations at run time and act accordingly. To this end, we test our proposed method on a number of test domains including the SAX Segmentation benchmark, which includes labelled test data from dense urban, rural and off-road driving domains. The proposed method consistently outperforms uncertainty estimation and Out-of-Distribution (OoD) techniques on this difficult benchmark.

Recent works on the pretraining for robot manipulation have demonstrated that representations learning from large human manipulation data can generalize well to new manipulation tasks and environments. However, these approaches mainly focus on human vision or natural language, neglecting tactile feedback. In this article, we make an attempt to explore how to pre-train a representation model for robotic manipulation using both human manipulation visual and tactile data. We develop a system for collecting visual and tactile data, featuring a cost-effective tactile glove to capture human tactile data and Hololens2 for capturing visual data. With this system, we collect a dataset of turning bottle caps. Furthermore, we introduce a novel visual-tactile fusion network and learning strategy M2VTP, with one key module to tokenize 20 sparse binary tactile signals sensing touch states for the learning of tactile context and the other key module applying the attention and mask mechanism to the interaction of visual and tactile tokens for visual-tactile representation learning. We utilize our dataset to pre-train the fusion model and embed the pre-trained model into a reinforcement learning framework for downstream tasks. Experimental results demonstrate that our pre-trained model significantly aids in learning manipulation skills. Compared to methods without pre-training, our approach achieves a success rate increase of over 60%. Additionally, when compared to current visual pre-training methods, our success rate exceeds them by more than 50%.

Point cloud is a popular and widely used geometric representation, which has attracted significant attention in 3D vision. However, the geometric variability of point cloud representations across different datasets can cause domain discrepancies, which hinder knowledge transfer and model generalization, resulting in degraded performance in target domain. In this paper, we present a novel approach to improve point cloud domain adaptation by employing masked representation learning in a self-supervised manner. Specifically, our method combines masked feature prediction and masked sample consistency to encode both local structure and global semantic information for learning invariant point cloud representation across domains. Moreover, to learn domain-specific representation and transfer knowledge from source to target, we propose prototype-calibrated self-training. By exploiting class-wise prototypes in the shared feature space, the soft pseudo labels can be adaptively denoised, which benefits the decision boundary learning in target domain. We conduct experiments on PointDA-10 and PointSegDA for 3D point cloud shape classification and semantic segmentation, respectively. The results demonstrate the effectiveness of our method and show that we can achieve the new state-of-the-art performance on point cloud domain adaptation.

Accurately perceiving and tracking instances over time is essential for the decision-making processes of autonomous agents interacting safely in dynamic environments. With this intention, we propose Mask4Former for the challenging task of 4D panoptic segmentation of LiDAR point clouds. Mask4Former is the first transformer-based approach unifying semantic instance segmentation and tracking of sparse and irregular sequences of 3D point clouds into a single joint model. Our model directly predicts semantic instances and their temporal associations without relying on hand-crafted non-learned association strategies such as probabilistic clustering or voting-based center prediction. Instead, Mask4Former introduces spatio-temporal instance queries that encode the semantic and geometric properties of each semantic tracklet in the sequence. In an in-depth study, we find that promoting spatially compact instance predictions is critical as spatiotemporal instance queries tend to merge multiple semantically similar instances, even if they are spatially distant. To this end, we regress 6-DOF bounding box parameters from spatiotemporal instance queries, which are used as an auxiliary task to foster spatially compact predictions. Mask4Former achieves a new state-of-the-art on the SemanticKITTI test set with a score of 68.4 LSTQ.

Given the image collection of an object, we aim at building a real-time image-based pose estimation method, which requires neither its CAD model nor hours of object-specific training. Recent NeRF-based methods provide a promising solution by directly optimizing the pose from pixel loss between rendered and target images. However, during inference, they require long converging time, and suffer from local minima, making them impractical for real-time robot applications. We aim at solving this problem by marrying image matching with NeRF. With 2D matches and depth rendered by NeRF, we directly solve the pose in one step by building 2D-3D correspondences between target and initial view, thus allowing for real-time prediction. Moreover, to improve the accuracy of 2D-3D correspondences, we propose a 3D consistent point mining strategy, which effectively discards unfaithful points reconstruted by NeRF. Moreover, current NeRF-based methods naively optimizing pixel loss fail at occluded images. Thus, we further propose a 2D matches based sampling strategy to preclude the occluded area. Experimental results on representative datasets prove that our method outperforms state-of-the-art methods, and improves inference efficiency by 90×, achieving real-time prediction at 6 FPS.

In mini-invasive surgery, the laparoscopic ultra-sound probe is visible in the laparoscopic image. We address the problem of estimating the probe pose with respect to the laparoscope without using markers and additional sensors. We propose the first method using a single standard laparoscopic monocular RGB image. It is robust, initialization-free and runs at 10 fps, thus forming a promising tool to improve robotic and augmented reality-based surgery.

Robots require a semantic understanding of their surroundings to operate in an efficient and explainable way in human environments. In the literature, there has been an extensive focus on object labeling and exhaustive scene graph generation; less effort has been focused on the task of purely identifying and mapping large semantic regions. The present work proposes a method for semantic region mapping via embodied navigation in indoor environments, generating a high-level representation of the knowledge of the agent. To enable region identification, the method uses a vision-to-language model to provide scene information for mapping. By projecting egocentric scene understanding into the global frame, the proposed method generates a semantic map as a distribution over possible region labels at each location. This mapping procedure is paired with a trained navigation policy to enable autonomous map generation. The proposed method significantly outperforms a variety of baselines, including an object-based system and a pretrained scene classifier, in experiments in a photorealistic simulator.

Learning a locomotion policy for quadruped robots has traditionally been constrained to a specific robot morphology, mass, and size. The learning process must usually be repeated for every new robot, where hyperparameters and reward function weights must be re-tuned to maximize performance for each new system. Alternatively, attempting to train a single policy to accommodate different robot sizes, while maintaining the same degrees of freedom (DoF) and morphology, requires either complex learning frameworks, or mass, inertia, and dimension randomization, which leads to prolonged training periods. In our study, we show that drawing inspiration from animal motor control allows us to effectively train a single locomotion policy capable of controlling a diverse range of quadruped robots. The robot differences encompass: a variable number of DoFs, (i.e. 12 or 16 joints), three distinct morphologies, a broad mass range spanning from 2 kg to 200 kg, and nominal standing heights ranging from 18 cm to 100 cm. Our policy modulates a representation of the Central Pattern Generator (CPG) in the spinal cord, effectively coordinating both frequencies and amplitudes of the CPG to produce rhythmic output (Rhythm Generation), which is then mapped to a Pattern Formation (PF) layer. Across different robots, the only varying component is the PF layer, which adjusts the scaling parameters for the stride height and length. Subsequently, we evaluate the sim-to-real transfer by testing the single policy on both the Unitree Go1 and A1 robots. Remarkably, we observe robust performance, even when adding a 15 kg load, equivalent to 125% of the A1 robot’s nominal mass.

For locomotion, is an arm on a legged robot a liability or an asset for locomotion? Biological systems evolved additional limbs beyond legs that facilitates postural control. This work shows how a manipulator can be an asset for legged locomotion at high speeds or under external perturbations, where the arm serves beyond manipulation. Since the system has 15 degrees of freedom (twelve for the legged robot and three for the arm), off-the-shelf reinforcement learning (RL) algorithms struggle to learn effective locomotion policies. Inspired by Bernstein’s neurophysiological theory of animal motor learning, we develop an incremental training procedure that initially freezes some degrees of freedom and gradually releases them, using behaviour cloning (BC) from an early learning procedure to guide optimization in later learning. Simulation experiments show that our policy increases the success rate by up to 61 percentage points over the baselines. Simulation and real robot experiments suggest that our policy learns to use the arm as a "tail" to initiate robot turning at high speeds and to stabilize the quadruped under external perturbations. Quantitatively, in simulation experiments, we cut the failure rate up to 43.6% during high-speed turning and up to 31.8% for quadruped under external forces compared to using a locked arm.