ICRA 2024 Papers

Embodied Instruction Following (EIF) is a task of planning a long sequence of sub-goals given high-level natural language instructions, such as "Rinse a slice of lettuce and place on the white table next to the fork". To successfully execute these long-term horizon tasks, we argue that an agent must consider its past, i.e., historical data, when making decisions in each step. Nevertheless, recent approaches in EIF often neglects the knowledge from historical data and also do not effectively utilize information across the modalities. To this end, we propose History-Aware Planning based on Fused Information(HAPFI), effectively leveraging the historical data from diverse modalities that agents collect while interacting with the environment. Specifically, HAPFI integrates multiple modalities, including historical RGB observations, bounding boxes, sub-goals, and high-level instructions, by effectively fusing modalities via our Mutually Attentive Fusion method. Through experiments with diverse comparisons, we show that an agent utilizing historical multi-modal information surpasses all the compared methods that neglect the historical data in terms of action planning capability, enabling the generation of well-informed action plans for the next step. Moreover, we provided qualitative evidence highlighting the significance of leveraging historical multi-modal data, particularly in scenarios where the agent encounters intermediate failures, showcasing its robust re-planning capabilities.

Robotic autonomous grasp requires the system to perform multiple functions such as gripper and robot control, making it a task with hybrid output nature. Existing methods based on closed-loop deep reinforcement learning rely on external models for termination evaluation. To achieve more effective grasp for novel objects, we propose a new autonomous grasp control scheme termed HAGrasp that considers the complete point cloud of the workspace. It integrates grasp pose estimation, end-effector pose evaluation, and motion planning of the robotic arm into a single model, enhancing the success rate while reducing computational load. We present a closed-loop grasp control system based on deep reinforcement learning. This control system can perform grasp tasks while dynamically adjusting to avoid end-effector collisions. The design of hybrid-action reinforcement learning module is trained with unified latent action space and further improve generalization, achieving real-time autonomous grasp control. Real robot experiments show that our method has 74.2% success rate for grasping 7 unseen objects. Comparative experiments show that the proposed HAGrasp outperforms open-loop baseline Contact-Graspnet in both success rate and inference time. It is demonstrated that with integrated multi-view input and sim-to-real training design, our method improves real-world applications of autonomous grasp.

Simultaneous Localization and Mapping (SLAM) has emerged as a prime autonomous mobile agent localization algorithm. Despite the global research effort to improve SLAM, its mapping component remains limited and serves little more than to satisfy the coupled localization problem. We present a collaborative 3D SLAM approach leveraging the power of augmented reality (AR). The system introduces a trio of diverse agents, each with its unique capability to become an active member in the mapping process: mobile robots, human operators, and AR head-mounted display (AR-HMD). A 3D complementary mapping pipeline is developed to utilize the built-in SLAM capabilities of the AR-HMD as shareable data. Our system aligns and merges the AR-HMD and the robot’s local map automatically, triggered by a human-dictated initial guess. The created merged map proves advantageous in scenarios where the robot is restricted from navigating in certain areas. To correct map imperfections resulting from problematic objects such as transparent or reflective surfaces, the fused map is overlayed onto the environment, and hand gestures are used to add or delete 3D map features in real-time. Our system is implemented in both a lab and a real industrial warehouse setup. The results show a significant improvement in the map quality and mapping duration.

In outdoor environments, Vision-and-Language Navigation (VLN) requires an agent to rely on multi-modal cues from real-world urban environments and natural language instructions. While existing outdoor VLN models predict actions using a combination of panorama and instruction features, this approach ignores objects in the environment and learns data bias to fail navigation. According to our preliminary findings, most instances of navigation failure in previous models were due to turning or stopping at the wrong place. In contrast, humans intuitively frequently use identifiable objects or store names as reference landmarks, ensuring accurate turns and stops, especially in unfamiliar places. To address this insight gap, we propose an Object-Attention VLN (OAVLN) model that helps the agent focus on relevant objects during training and understand the environment better. Our model outperforms previous methods in all evaluation metrics under both seen and unseen scenarios on two existing benchmark datasets, Touchdown and map2seq.

Safe Reinforcement Learning (RL) aims to find a policy that achieves high rewards while satisfying cost constraints. When learning from scratch, safe RL agents tend to be overly conservative, which impedes exploration and restrains the overall performance. In many realistic tasks, e.g. autonomous driving, large-scale expert demonstration data are available. We argue that extracting expert policy from offline data to guide online exploration is a promising solution to mitigate the conserveness issue. Large-capacity models, e.g. decision transformers (DT), have been proven to be competent in offline policy learning. However, data collected in realworld scenarios rarely contain dangerous cases (e.g., collisions), which makes it prohibitive for the policies to learn safety concepts. Besides, these bulk policy networks cannot meet the computation speed requirements at inference time on real-world tasks such as autonomous driving. To this end, we propose Guided Online Distillation (GOLD), an offline-to-online safe RL framework. GOLD distills an offline DT policy into a lightweight policy network through guided online safe RL training, which outperforms both the offline DT policy and online safe RL algorithms. Experiments in both benchmark safe RL tasks and real-world driving tasks based on the Waymo Open Motion Dataset (WOMD) [1] demonstrate that GOLD can successfully distill lightweight policies and solve decision-making problems in challenging safety-critical scenarios.

Deep reinforcement learning can enable robots to autonomously acquire complex behaviors such as legged locomotion. However, RL in the real world is complicated by constraints on efficiency, safety, and overall training stability, which limits its practical applicability. We present APRL, a policy regularization framework that modulates the robot’s exploration throughout training, striking a balance between flexible improvement potential and focused, efficient exploration. APRL enables a quadrupedal robot to efficiently learn to walk entirely in the real world within minutes and continue to improve with more training where prior work saturates in performance. We demonstrate that continued training with APRL results in a policy that is substantially more capable of navigating challenging situations and adapts to changes in dynamics. Videos and code to reproduce our results are available at: https://sites.google.com/berkeley.edu/aprl

This work explores a novel approach to empowering robots with visual perception capabilities using textual descriptions. Our approach involves the integration of GPT-4 with dense captioning, enabling robots to perceive and interpret the visual world through detailed text-based descriptions. To assess both user experience and the technical feasibility of this approach, experiments were conducted with human participants interacting with a Pepper robot equipped with visual capabilities. The results affirm the viability of the proposed approach, allowing to perform vision-based conversations effectively, despite processing time limitations.

We introduce Ground-Fusion, a low-cost sensor fusion simultaneous localization and mapping (SLAM) system for ground vehicles. Our system features efficient initialization, effective sensor anomaly detection and handling, real-time dense color mapping, and robust localization in diverse environments. We tightly integrate RGB-D images, inertial measurements, wheel odometer and GNSS signals within a factor graph to achieve accurate and reliable localization both indoors and outdoors. To ensure successful initialization, we propose an efficient strategy that comprises three different methods: stationary, visual, and dynamic, tailored to handle diverse cases. Furthermore, we develop mechanisms to detect sensor anomalies and degradation, handling them adeptly to maintain system accuracy. Our experimental results on both public and self-collected datasets demonstrate that Ground-Fusion outperforms existing low-cost SLAM systems in corner cases. We release the code and datasets at https://github.com/SJTU-ViSYS/Ground-Fusion.

Foundation models such as ChatGPT have made significant strides in robotic tasks due to their universal representation of real-world domains. In this paper, we leverage foundation models to tackle grasp detection, a persistent challenge in robotics with broad industrial applications. Despite numerous grasp datasets, their object diversity remains limited compared to real-world figures. Fortunately, foundation models possess an extensive repository of real-world knowledge, including objects we encounter in our daily lives. As a consequence, a promising solution to the limited representation in previous grasp datasets is to harness the universal knowledge embedded in these foundation models. We present Grasp-Anything, a new large-scale grasp dataset synthesized from foundation models to implement this solution. Grasp-Anything excels in diversity and magnitude, boasting 1M samples with text descriptions and more than 3M objects, surpassing prior datasets. Empirically, we show that Grasp-Anything successfully facilitates zero-shot grasp detection on vision-based tasks and real-world robotic experiments. Our dataset and code are available at https://airvlab.github.io/grasp-anything/.

In multi-object stacking scenarios, exploring the relationships among objects and determining the correct sequence of operations are crucial for robotic manipulation. However, previous algorithms inefficiently combine global and local information, often focusing solely on the local features of objects or the interactions of object features at a global level. This approach leads to imbalanced distribution of features and the generation of redundant or missing relationships in complex scenes, such as multi-object stacking and partial occlusion. To address this issue, we have developed a grasp manipulation relationship detection algorithm called Graph Sampling Aggregation Network for Visual Manipulation Relationship Detection (GSAGED). This algorithm assists robots in detecting targets in complex scenes and determining the appropriate grasping order. Firstly, the Positional Encoding Module in GSAGED enhances object feature information by considering global contexts. Secondly, the Graph Sampling Aggregation method effectively integrates global and local information, relieving imbalanced distribution of features. Finally, we applied the developed algorithm to a physical robot for grasping. Experimental results on the Visual Manipulation Relationship Dataset (VMRD) and the large-scale relational grasp dataset named REGRAD demonstrate that our method significantly improves the accuracy of relationship detection in complex scenes and exhibits robust generalization capabilities in real-world applications.

Interactive grasping from clutter, akin to human dexterity, is one of the longest-standing problems in robot learning. Challenges stem from the intricacies of visual perception, the demand for precise motor skills, and the complex interplay between the two. In this work, we present Teacher-Augmented Policy Gradient (TAPG), a novel two-stage learning framework that synergizes reinforcement learning and policy distillation. After training a teacher policy to master the motor control based on object pose information, TAPG facilitates guided, yet adaptive, learning of a sensorimotor policy, based on object segmentation. We zero-shot transfer from simulation to a real robot by using Segment Anything Model for promptable object segmentation. Our trained policies adeptly grasp a wide variety of objects from cluttered scenarios in simulation and the real world based on human-understandable prompts. Furthermore, we show robust zero-shot transfer to novel objects. Videos of our experiments are available at https://maltemosbach.github.io/grasp_anything.

We aim to solve the problem of data-driven collision-distance estimation given 3-dimensional (3D) geometries. Conventional algorithms suffer from low accuracy due to their reliance on limited representations, such as point clouds. In contrast, our previous graph-based model, GraphDistNet, achieves high accuracy using edge information but incurs higher message-passing costs with growing graph size, limiting its applicability to 3D geometries. To overcome these challenges, we propose GDN-R, a novel 3D graph-based estimation network. GDN-R employs a layer-wise probabilistic graph-rewiring algorithm leveraging the differentiable Gumbel-top-K relaxation. Our method accurately infers minimum distances through iterative graph rewiring and updating relevant embeddings. The probabilistic rewiring enables fast and robust embedding with respect to unforeseen categories of geometries. Through 41, 412 random benchmark tasks with 150 pairs of 3D objects, we show GDN-R outperforms state-of-the-art baseline methods in terms of accuracy and generalizability. We also show that the proposed rewiring improves the update performance reducing the size of the estimation model. We finally show its batch prediction and auto-differentiation capabilities for trajectory optimization in both simulated and real-world scenarios.

One goal of dexterous robotic grasping is to allow robots to handle objects with the same level of flexibility and adaptability as humans. However, it remains a challenging task to generate an optimal grasping strategy for dexterous hands, especially when it comes to delicate manipulation and accurate adjustment the desired grasping poses for objects of varying shapes and sizes. In this paper, we propose a novel dexterous grasp generation scheme called GrainGrasp that provides fine-grained contact guidance for each fingertip. In particular, we employ a generative model to predict separate contact maps for each fingertip on the object point cloud, effectively capturing the specifics of finger-object interactions. In addition, we develop a new dexterous grasping optimization algorithm that solely relies on the point cloud as input, eliminating the necessity for complete mesh information of the object. By leveraging the contact maps of different fingertips, the proposed optimization algorithm can generate precise and determinable strategies for human-like object grasping. Experimental results confirm the efficiency of the proposed scheme. Our code is available at https://github.com/wmtlab/GrainGrasp.

Robots are increasingly used in tomato greenhouses to automate labour-intensive tasks such as selective harvesting and de-leafing. To perform these tasks, robots must be able to accurately and efficiently perceive the plant nodes that need to be cut, despite the high levels of occlusion from other plant parts. We formulate this problem as a local next-best-view (NBV) planning task where the robot has to plan an efficient set of camera viewpoints to overcome occlusion and improve the quality of perception. Our formulation focuses on quickly improving the perception accuracy of a single target node to maximise its chances of being cut. Previous methods of NBV planning mostly focused on global view planning and used random sampling of candidate viewpoints for exploration, which could suffer from high computational costs, ineffective view selection due to poor candidates, or non-smooth trajectories due to inefficient sampling. We propose a gradient-based NBV planner using differentiable ray sampling, which directly estimates the local gradient direction for viewpoint planning to overcome occlusion and improve perception. Through simulation experiments, we showed that our planner can handle occlusions and improve the 3D reconstruction and position estimation of nodes equally well as a sampling-based NBV planner, while taking ten times less computation and generating 28% more efficient trajectories.

Imitation learning mitigates the resource-intensive nature of learning policies from scratch by mimicking expert behavior. While existing methods can accurately replicate expert demonstrations, they often exhibit unpredictability in unexplored regions of the state space, thereby raising major safety concerns when facing perturbations. We propose SNDS, an imitation learning approach aimed at efficient training of scalable neural policies while formally ensuring global stability. SNDS leverages a neural architecture that enables the joint training of the policy and its associated Lyapunov candidate to ensure global stability throughout the learning process. We validate our approach through extensive simulations and deploy the trained policies on a real-world manipulator arm. The results confirm SNDS’s ability to address instability, accuracy, and computational intensity challenges highlighted in the literature, positioning it as a promising solution for scalable and stable policy learning in complex environments.

Visual localization is an important sub-task in SfM and visual SLAM that involves estimating a 6-DoF camera pose for an input query image relative to a given 3D model of the environment. The most accurate approach is a hierarchical one that splits the task into two stages: image retrieval and camera pose estimation. Each stage requires different image features, with global features compactly encoding holistic image information for the first stage and local features encoding the appearance around salient image points for the second stage. While existing methods use independent networks to extract these features, one for global and one for local, this strategy is suboptimal in terms of computational efficiency. In this paper, we propose a novel approach that achieves state-of-the-art inference accuracy with significantly improved efficiency. Our approach’s core component is SuperGF, a network that aggregates local features optimized for camera pose estimation to create a global feature that enables precise image retrieval. Through extensive experiments on the standard benchmark tests, we demonstrate that the method offers a better trade-off between accuracy and computational cost.

This paper researches a novel global terminal sliding mode control(GTSMC) on a tethered satellites system(TSS) under outer disturbances, and the effect of PI/PD compensation in restraining chattering on sliding surface is appended. By taking advantage of the finite-time convergence of traditional terminal sliding surface, the sliding surface with global and terminal sliding motion is proposed, and the convergent time by GTSMC is qualitatively evaluated by the sliding surface. Then the integral/derivative function of the low-pass filtered switching control is appended in GTSMC. By virtue of the accuracy of integral and the damping of derivative, respectively, the persisting on sliding surface is eliminated, such that the chattering effect of the controlled system on the surface is restrained consequently. Finally, simulations of the proposed control on TSS is shown to validate the theoretical analyses.

Multi-view stereo omnidirectional distance estimation usually needs to build a cost volume with many hypothetical distance candidates. The cost volume building process is often computationally heavy considering the limited resources a mobile robot has. We propose a new geometry-informed way of distance candidates selection method which enables the use of a very small number of candidates and reduces the computational cost. We demonstrate the use of the geometry-informed candidates in a set of model variants. We find that by adjusting the candidates during robot deployment, our geometry-informed distance candidates also improve a pre-trained model’s accuracy if the extrinsics or the number of cameras changes. Without any re-training or fine-tuning, our models outperform models trained with evenly distributed distance candidates. Models are also released as hardware-accelerated versions with a new dedicated large-scale dataset. The project page, code, and dataset can be found at https://theairlab.org/gicandidates/.

This work focuses on the agile transportation of liquids with robotic manipulators. In contrast to existing methods that are either computationally heavy, system/container specific or dependant on a singularity-prone pendulum model, we present a real-time slosh-free tracking technique. This method solely requires the reference trajectory and the robot’s kinematic constraints to output kinematically feasible joint space commands. The crucial element underlying this approach consists on mimicking the end-effector’s motion through a virtual quadrotor, which is inherently slosh-free and differentially flat, thereby allowing us to calculate a slosh-free reference orientation. Through the utilization of a cascaded proportional-derivative (PD) controller, this slosh-free reference is transformed into task space acceleration commands, which, following the resolution of a Quadratic Program (QP) based on Resolved Acceleration Control (RAC), are translated into a feasible joint configuration. The validity of the proposed approach is demonstrated by simulated and real-world experiments on a 7 DoF Franka Emika Panda robot.