IROS 2024 论文列表

Being widespread in human motor behavior, dynamic movements demonstrate higher efficiency and greater capacity to address a broader range of skill domains compared to their quasi-static counterparts. Among the frequently studied dynamic manipulation problems, robotic juggling tasks stand out due to their inherent ability to scale their difficulty levels to arbitrary extents, making them an excellent subject for investigation. In this study, we explore juggling patterns with mixed throw heights, following the vanilla siteswap juggling notation, which jugglers widely adopted to describe toss juggling patterns. This requires extending our previous analysis of the simpler cascade juggling task by a throw-height sequence planner and further constraints on the end effector trajectory. These are not necessary for cascade patterns but are vital to achieving patterns with mixed throw heights. Using a simulated environment, we demonstrate successful juggling of most common 3-9 ball siteswap patterns up to 9 ball height, transitions between these patterns, and random sequences covering all possible vanilla siteswap patterns with throws between 2 and 9 ball height. https://kai-ploeger.com/beyond-cascades

Learning from Demonstration (LfD) empowers novice users to teach robots daily life tasks without writing sophisticated code, thereby promoting the democratization of robotics. However, novice users often provide sub-optimal demonstrations, which can potentially impact the robot’s ability to efficiently learn and execute the tasks. Prior research has assessed the quality of demonstrations by evaluating the robot’s task performance; however, the approach remains insufficient to qualify individual demonstrations, leaving the reason for classifying demonstrations as high- or low-quality unknown. Therefore, this simulation-based study aims to quantify the quality of individual demonstration at each step by incorporating motion-related quality features such as manipulability and joint-space jerk. To assess the efficacy of these features, we initially evaluated the given demonstrations—taking into account each quality feature—to rank them from high- to low-quality. Subsequently, we investigated the impact of demonstration’s quality on task performance and the quality of task execution. In this pursuit, we trained a series of LfD models for distinct manipulation tasks: cube lifting and pick-and-place of soda can. Our results illustrate a strong correlation between ranked demonstrations and the quality of task execution. Interestingly, we observed that the quality features have a significant impact on task performance, particularly when the provided demonstrations exhibit diversity in terms of quality. Overall, this analysis enables quantifying the quality of individual demonstrations based on motion-related quality features, thus improving learning from demonstration.

Advancements in Industry 4.0 demand sophisticated solutions for automatic robotic assembly sequence planning (RASP), capable of handling the diversity and complexity of modern manufacturing tasks. One approach to RASP is Assembly-by-Disassembly (AbD). It first searches for a disassembly sequence that is then inverted to obtain an assembly sequence. One of the challenges of AbD, however, is the exponential number of potential assembly sequences for any given assembly. To mitigate this challenge, we propose to transfer knowledge obtained during previous planning attempts. Specifically, we present an approach that combines Monte Carlo Tree Search (MCTS) with deep Q-learning to optimize the total length of robotic assembly paths. We use a graph-based representation of disassembly states in combination with a graph neural network to learn the Q-function. We further discuss a principled approach to generate 3D assemblies out of aluminium profiles that a single robot manipulator can assemble. With this approach, we generated two datasets consisting of 14 assemblies with 21 removable parts and 7 assemblies with 30 removable parts. Using leave-one-out cross-validation, we were able to demonstrate how our approach outperformed an unmodified MCTS. Moreover, we successfully transferred knowledge between datasets.

In the field of robotics, event-based cameras are emerging as a promising low-power alternative to traditional frame-based cameras for capturing high-speed motion and high dynamic range scenes. This is due to their sparse and asynchronous event outputs. Spiking Neural Networks (SNNs) with their asynchronous event-driven compute, show great potential for extracting the spatio-temporal features from these event streams. In contrast, the standard Analog Neural Networks (ANNs1) fail to process event data effectively. However, training SNNs is difficult due to additional trainable parameters (thresholds and leaks), vanishing spikes at deeper layers, and a non-differentiable binary activation function. Furthermore, an additional data structure, "membrane potential", responsible for keeping track of temporal information, must be fetched and updated at every timestep in SNNs. To overcome these challenges, we propose a novel SNN-ANN hybrid architecture that combines the strengths of both. Specifically, we leverage the asynchronous compute capabilities of SNN layers to effectively extract the input temporal information. Concurrently, the ANN layers facilitate training and efficient hardware deployment on traditional machine learning hardware such as GPUs. We provide extensive experimental analysis for assigning each layer to be spiking or analog, leading to a network configuration optimized for performance and ease of training. We evaluate our hybrid architecture for optical flow estimation on DSEC-flow and Multi-Vehicle Stereo Event-Camera (MVSEC) datasets. On the DSEC-flow dataset, the hybrid SNN-ANN architecture achieves a 40% reduction in average endpoint error (AEE) with 22% lower energy consumption compared to Full-SNN, and 48% lower AEE compared to Full-ANN, while maintaining comparable energy usage.

Reinforcement learning (RL) policies are prone to high-frequency oscillations, especially undesirable when deploying to hardware in the real-world. In this paper, we identify, categorize, and compare methods from the literature that aim to mitigate high-frequency oscillations in deep RL. We define two broad classes: loss regularization and architectural methods. At their core, these methods incentivize learning a smooth mapping, such that nearby states in the input space produce nearby actions in the output space. We present benchmarks in terms of policy performance and control smoothness on traditional RL environments from the Gymnasium and a complex manipulation task, as well as three robotics locomotion tasks that include deployment and evaluation with real-world hardware. Finally, we also propose hybrid methods that combine elements from both loss regularization and architectural methods. We find that the best-performing hybrid outperforms other methods, and improves control smoothness by 26.8% over the baseline, with a worst-case performance degradation of just 2.8%.

Recent research on mobile robot navigation has focused on socially aware navigation in crowded environments. However, existing methods do not adequately account for human–robot interactions and demand accurate location information from omnidirectional sensors, rendering them unsuitable for practical applications. In response to this need, this study introduces a novel algorithm, BNBRL+, predicated on the partially observable Markov decision process framework to assess risks in unobservable areas and formulate movement strategies under uncertainty. BNBRL+ consolidates belief algorithms with Bayesian neural networks to probabilistically infer beliefs based on the positional data of humans. It further integrates the interactions between the robot, humans, and inferred beliefs to determine the navigation paths, thereby facilitating socially aware navigation. Through experiments in various risk-laden scenarios, this study validates the effectiveness of BNBRL+ in navigating crowded environments with blind spots. The model’s ability to navigate effectively in spaces with limited visibility and avoid obstacles dynamically can significantly improve the safety and reliability of autonomous vehicles. The complement source code can be accessed here: https://github.com/JinnnK/BNBRLplus.

Language-conditioned Robotic manipulation demonstrates great potential in tackling various tasks. However, the generalization ability of this technique to unseen commands remains a challenge. Moreover, existing methods suffer from the burdensome overhead of data collection costs. Nowadays, Large language models (LLMs) have demonstrated impressive natural language understanding capabilities. In this work, we propose a novel scheme called Behavior-Actor(BehAct), which leverages the power of LLM to decompose language commands into executable behaviors in Retrieval-Augmented Generation(RAG) manner. A End-to-End actor is then trained to execute these identified behaviors. BehAct’s LLM acts as a "brain", while the actor acts as a "hand". A single actor model is trained from scratch on 11 real-world tasks, 40 behaviors using 276 demonstrations, only 7 for each behavior in average. We achieve a 68% average success rate on seen commands, which aligns comparably with recent works. Moreover, BehAct exhibits an impressive 45% average success rate on unseen commands, doubling the performance of the baseline approach. In the BehAct system, LLM-agnostic design enables flexibility in leveraging advanced LLMs without necessitating fine-tuning. Our code has been made publicly available here.

In this article, we present a reactive multi-agent coordination architecture for the management of material flows between production/pickup stages and delivery/drop-off stages, in scenarios such as underground mines and automated factory floors. The pickup and delivery stages are modelled as variable task queues, with no a priori information about the inflow into the production queues. The proposed solution coordinates the movement of a group of mobile agents operating between the two stages in a reactive and scalable manner, so that the material is transported from multiple production queues to multiple delivery queues in a balanced/equalized manner. In such a scenario, centralized planners suffer from low reactivity and poor scaling, as the number of agents and number of queues increases. To overcome this problem, we propose a decentralized approach comprising of two separate auction-based task distribution systems for the production and delivery stages, along with behavior-tree based management of agent autonomy and task bidding. Each auction system tracks the length of production/delivery queues and solves the optimal task assignment, based on the bids submitted by the agents. The agents participate in one of the two auction systems at any given time, based on the status of the behavior tree executing the two-stage tasks. We analytically show that the proposed decentralized auctioning approach along with agent autonomy and bidding managed by behavior trees, offers better scalability and reactiveness compared to the centralized approach. The proposed methodology is experimentally validated in a lab environment, in three illustrative material flow management scenarios, using TurtleBot3 robots as agents.

Sample efficient learning of manipulation skills poses a major challenge in robotics. While recent approaches demonstrate impressive advances in the type of task that can be addressed and the sensing modalities that can be incorporated, they still require large amounts of training data. Especially with regard to learning actions on robots in the real world, this poses a major problem due to the high costs associated with both demonstrations and real-world robot interactions. To address this challenge, we introduce BOpt-GMM, a hybrid approach that combines imitation learning with own experience collection. We first learn a skill model as a dynamical system encoded in a Gaussian Mixture Model from a few demonstrations. We then improve this model with Bayesian optimization building on a small number of autonomous skill executions in a sparse reward setting. We demonstrate the sample efficiency of our approach on multiple complex manipulation skills in both simulations and real-world experiments. Furthermore, we make the code and pre-trained models publicly available at http://bopt-gmm.cs.uni-freiburg.de.

Mapping people dynamics is a crucial skill for robots, because it enables them to coexist in human-inhabited environments. However, learning a model of people dynamics is a time consuming process which requires observation of large amount of people moving in an environment. Moreover, approaches for mapping dynamics are unable to transfer the learned models across environments: each model is only able to describe the dynamics of the environment it has been built in. However, the impact of architectural geometry on people’s movement can be used to anticipate their patterns of dynamics, and recent work has looked into learning maps of dynamics from occupancy. So far however, approaches based on trajectories and those based on geometry have not been combined. In this work we propose a novel Bayesian approach to learn people dynamics able to combine knowledge about the environment geometry with observations from human trajectories. An occupancy-based deep prior is used to build an initial transition model without requiring any observations of pedestrian; the model is then updated when observations become available using Bayesian inference. We demonstrate the ability of our model to increase data efficiency and to generalize across real large-scale environments, which is unprecedented for maps of dynamics.

Domain randomization (DR), which entails training a policy with randomized dynamics, has proven to be a simple yet effective algorithm for reducing the gap between simulation and the real world. However, DR often requires careful tuning of randomization parameters. Methods like Bayesian Domain Randomization (Bayesian DR) and Active Domain Randomization (Adaptive DR) address this issue by automating parameter range selection using real-world experience. While effective, these algorithms often require long computation time, as a new policy is trained from scratch every iteration. In this work, we propose Adaptive Bayesian Domain Randomization via Strategic Fine-tuning (BayRnTune), which inherits the spirit of BayRn but aims to significantly accelerate the learning processes by fine-tuning from previously learned policy. This idea leads to a critical question: which previous policy should we use as a prior during fine-tuning? We investigated four different fine-tuning strategies and compared them against baseline algorithms in five simulated environments, ranging from simple benchmark tasks to more complex legged robot environments. Our analysis demonstrates that our method yields better rewards in the same amount of timesteps compared to vanilla domain randomization or Bayesian DR.

The production of wind energy is a crucial part of sustainable development and reducing the reliance on fossil fuels. Maintaining the integrity of wind turbines to produce this energy is a costly and time-consuming task requiring repeated inspection and maintenance. While autonomous drones have proven to make this process more efficient, the algorithms for detecting anomalies to prevent catastrophic damage to turbine blades have fallen behind due to some dangerous defects, such as hairline cracks, being barely-visible. Existing datasets and literature are lacking and tend towards detecting obvious and visible defects in addition to not being geographically diverse. In this paper we introduce a novel and diverse dataset of barelyvisible hairline cracks collected from numerous wind turbine inspections. To prove the efficacy of our dataset, we detail our end-to-end deployed turbine crack detection pipeline from the image acquisition stage to the use of predictions in providing automated maintenance recommendations to extend the life and efficiency of wind turbines.

To enhance accuracy of robot state estimation, active sensing (or perception-aware) methods seek trajectories that maximize the information gathered by the sensors. To this aim, one possibility is to seek trajectories that minimize the estimation error covariance matrix output by an extended Kalman filter (EKF), w.r.t. its control inputs over a given horizon. However, this is computationally demanding. In this article, we derive novel backpropagation analytical formulas for the derivatives of the covariance matrices of an EKF w.r.t. all its inputs. We then leverage the obtained analytical gradients as an enabling technology to derive perception-aware optimal motion plans. Simulations validate the approach, showcasing improvements in execution time, notably over PyTorch’s automatic differentiation. Experimental results on a real vehicle also support the method.

In many applications, a mobile manipulator robot is required to grasp a set of objects distributed in space. This may not be feasible from a single base pose and the robot must plan the sequence of base poses for grasping all objects, minimizing the total navigation and grasping time. This is a Combinatorial Optimization problem that can be solved using exact methods, which provide optimal solutions but are computationally expensive, or approximate methods, which offer computationally efficient but sub-optimal solutions. Recent studies have shown that learning-based methods can solve Combinatorial Optimization problems, providing near-optimal and computationally efficient solutions.In this work, we present BaSeNet - a learning-based approach to plan the sequence of base poses for the robot to grasp all the objects in the scene. We propose a Reinforcement Learning based solution that learns the base poses for grasping individual objects and the sequence in which the objects should be grasped to minimize the total navigation and grasping costs using Layered Learning. As the problem has a varying number of states and actions, we represent states and actions as a graph and use Graph Neural Networks for learning. We show that the proposed method can produce comparable solutions to exact and approximate methods with significantly less computation time. The code and Reinforcement Learning environments will be made available on the project webpage*.

We present a new approach to robot hand design specifically suited to enable robot learning methods and daily tasks in human environments. We introduce BaRiFlex, an innovative gripper design that alleviates the issues caused by unexpected contact and collisions during robot learning, offering collision mitigation, grasping versatility, task versatility, and simplicity to the learning processes. This achievement is enabled by the incorporation of low-inertia actuators, providing high Back-drivability, and the strategic combination of Rigid and Flexible materials which enhances versatility and the gripper’s resilience against unpredicted collisions. Furthermore, the integration of flexible Fin-Ray and rigid linkages allows the gripper to execute compliant grasping and precise pinching. We conducted rigorous performance tests to characterize the novel gripper’s compliance, durability, grasping and task versatility, and precision. We also integrated the BaRiFlex with a 7 Degree of Freedom (DoF) Franka Emika’s Panda robotic arm to evaluate its capacity to support a trial-and-error (reinforcement learning) training procedure. The results of our experimental study are then compared to those obtained using the original rigid Franka Hand and a reference Fin-Ray soft gripper, demonstrating the superior capabilities and advantages of our developed gripper system. More information and videos at https://robin-lab.cs.utexas.edu/bariflex

This paper presents a novel approach to generating behavior trees for robots using lightweight large language models (LLMs) with a maximum of 7 billion parameters. The study demonstrates that it is possible to achieve satisfying results with compact LLMs when fine-tuned on a specific dataset. The key contributions of this research include the creation of a fine-tuning dataset based on existing behavior trees using GPT-3.5 and a comprehensive comparison of multiple LLMs (namely llama2, llama-chat, and code-llama) across nine distinct tasks. To be thorough, we evaluated the generated behavior trees using static syntactical analysis, a validation system, a simulated environment, and a real robot. Furthermore, this work opens the possibility of deploying such solutions directly on the robot, enhancing its practical applicability. Findings from this study demonstrate the potential of LLMs with a limited number of parameters in generating effective and efficient robot behaviors.

In logistics, the ability to quickly compute and execute pick-and-place motions from bins is critical to increasing productivity. We present Bin-Optimized Motion Planning (BOMP), a motion planning framework that plans arm motions for a six-axis industrial robot with a long-nosed suction tool to remove boxes from deep bins. BOMP considers robot arm kinematics, actuation limits, the dimensions of a grasped box, and a varying height map of a bin environment to rapidly generate time-optimized, jerk-limited, and collision-free trajectories. The optimization is warm-started using a deep neural network trained offline in simulation with 25,000 scenes and corresponding trajectories. Experiments with 96 simulated and 15 physical environments suggest that BOMP generates collision-free trajectories that are up to 58% faster than baseline sampling-based planners and up to 36% faster than an industry-standard Up-Over-Down algorithm, which has an extremely low 15% success rate in this context. BOMP also generates jerk-limited trajectories while baselines do not. Website: https://sites.google.com/berkeley.edu/bomp.

We introduce BEVRender, a novel learning-based approach for the localization of ground vehicles in Global Navigation Satellite System (GNSS)-denied off-road scenarios. These environments are typically challenging for conventional vision-based state estimation due to the lack of distinct visual landmarks and the instability of vehicle poses. To address this, BEVRender generates high-quality local bird’s-eye-view (BEV) images of the local terrain. Subsequently, these images are aligned with a georeferenced aerial map through template matching to achieve accurate cross-view registration. Our approach overcomes the inherent limitations of visual inertial odometry systems and the substantial storage requirements of image-retrieval localization strategies, which are susceptible to drift and scalability issues, respectively. Extensive experimentation validates BEVRender’s advancement over existing GNSS-denied visual localization methods, demonstrating notable enhancements in both localization accuracy and update frequency.

Ground to aerial matching is a crucial and challenging task in outdoor robotics, particularly when GPS is absent or unreliable. Structures like buildings or large dense forests create interference, requiring GNSS replacements for global positioning estimates. The true difficulty lies in reconciling the perspective difference between the ground and air images for acceptable localization.Taking inspiration from the autonomous driving community, we propose a novel framework for synthesizing a birds-eye-view (BEV) scene representation to match and localize against an aerial map in off-road environments. We leverage contrastive learning with domain specific hard negative mining to train a network to learn similar representations between the synthesized BEV and the aerial map.During inference, BEVLoc guides the identification of the most probable locations within the aerial map through a coarse-to-fine matching strategy. Our results demonstrate promising initial outcomes in extremely difficult forest environments with limited semantic diversity. We analyze our model’s performance for coarse and fine matching, assessing both the raw matching capability of our model and its performance as a GNSS replacement.Our work delves into off-road map localization while establishing a foundational baseline for future developments in localization. Our code is available at: https://github.com/rpl-cmu/bevloc