ICRA 2024 论文列表

Since autonomous driving systems usually face dynamic and ever-changing environments, continual test-time adaptation (CTTA) has been proposed as a strategy for transferring deployed models to continually changing target domains. However, the pursuit of long-term adaptation often introduces catastrophic forgetting and error accumulation problems, which impede the practical implementation of CTTA in the real world. Recently, existing CTTA methods mainly focus on utilizing a majority of parameters to fit target domain knowledge through self-training. Unfortunately, these approaches often amplify the challenge of error accumulation due to noisy pseudo-labels, and pose practical limitations stemming from the heavy computational costs associated with entire model updates. In this paper, we propose a distribution-aware tuning (DAT) method to make the semantic segmentation CTTA efficient and practical in real-world applications. DAT adaptively selects and updates two small groups of trainable parameters based on data distribution during the continual adaptation process, including domain-specific parameters (DSP) and task-relevant parameters (TRP). Specifically, DSP exhibits sensitivity to outputs with substantial distribution shifts, effectively mitigating the problem of error accumulation. In contrast, TRP are allocated to positions that are responsive to outputs with minor distribution shifts, which are fine-tuned to avoid the catastrophic forgetting problem. In addition, since CTTA is a temporal task, we introduce the Parameter Accumulation Update (PAU) strategy to collect the updated DSP and TRP in target domain sequences. We conducted extensive experiments on two widely-used semantic segmentation CTTA benchmarks, achieving competitive performance and efficiency compared to previous state-of-the-art methods.

In multi-robot informative path planning the problem is to find a route for each robot in a team to visit a set of locations that can provide the most useful data to reconstruct an unknown scalar field. In the budgeted version, each robot is subject to a travel budget limiting the distance it can travel. Our interest in this problem is motivated by applications in precision agriculture, where robots are used to collect measurements to estimate domain-relevant scalar parameters such as soil moisture or nitrates concentrations. In this paper, we propose an online, distributed multi-robot sampling algorithm based on Monte Carlo Tree Search (MCTS) where each robot iteratively selects the next sampling location through communication with other robots and considering its remaining budget.We evaluate our proposed method for varying team sizes and in different environments, and we compare our solution with four different baseline methods. Our experiments show that our solution outperforms the baselines when the budget is tight by collecting measurements leading to smaller reconstruction errors.

In this work, we propose a framework for the distributed implementation of Quadratic Programs-based controllers, building upon and rectifying a significant limitation in a previously presented approach. The proposed framework is primarily motivated by the distributed implementation of Control Barrier Functions (CBFs), whose primary objective is to make minimal adjustments to a nominal controller while ensuring constraint satisfaction. By improving over some limitations in the current state-of-the-art, we are able to apply distributed CBFs to the problem of global connectivity maintenance in presence of communication and sensing constraints. Specifically, we consider the problem of preserving connectivity for a group of quadrotors with onboard sensors under distance and field of view constraints. Leveraging distributed control barrier functions, our approach maintains global graph connectivity while optimizing the performance of the desired task. Numerical simulations validate its effectiveness.

Today’s robot policies exhibit subpar performance when faced with the challenge of generalizing to novel environments. Human corrective feedback is a crucial form of guidance to enable such generalization. However, adapting to and learning from online human corrections is a non-trivial endeavor: not only do robots need to remember human feedback over time to retrieve the right information in new settings and reduce the intervention rate, but also they would need to be able to respond to feedback that can be arbitrary corrections about high-level human preferences to low-level adjustments to skill parameters. In this work, we present Distillation and Retrieval of Online Corrections (DROC), a large language model (LLM)-based system that can respond to arbitrary forms of language feedback, distill generalizable knowledge from corrections, and retrieve relevant past experiences based on textual and visual similarity for improving performance in novel settings. DROC is able to respond to a sequence of online language corrections that address failures in both high-level task plans and low-level skill primitives. We demonstrate that DROC effectively distills the relevant information from the sequence of online corrections in a knowledge base and retrieves that knowledge in settings with new task or object instances. DROC outperforms other techniques that directly generate robot code via LLMs [1] by using only half of the total number of corrections needed in the first round and requires little to no corrections after two iterations. We show further results and videos on our project website: https://sites.google.com/stanford.edu/droc.

In recent years, numerous real-time stereo matching methods have been introduced, but they often lack accuracy. These methods attempt to improve accuracy by introducing new modules or integrating traditional methods. However, the improvements are only modest. In this paper, we propose a novel strategy by incorporating knowledge distillation and model pruning to overcome the inherent trade-off between speed and accuracy. As a result, we obtained a model that maintains real-time performance while delivering high accuracy on edge devices. Our proposed method involves three key steps. Firstly, we review state-of-the-art methods and design our lightweight model by removing redundant modules from those efficient models through a comparison of their contributions. Next, we leverage the efficient model as the teacher to distill knowledge into the lightweight model. Finally, we systematically prune the lightweight model to obtain the final model. Through extensive experiments conducted on two widely-used benchmarks, Sceneflow and KITTI, we perform ablation studies to analyze the effectiveness of each module and present our state-of-the-art results.

Visual language navigation (VLN) is an embodied task demanding a wide range of skills encompassing understanding, perception, and planning. For such a multifaceted challenge, previous VLN methods totally rely on one model’s own thinking to make predictions within one round. However, existing models, even the most advanced large language model GPT4, still struggle with dealing with multiple tasks by single-round self-thinking. In this work, drawing inspiration from the expert consultation meeting, we introduce a novel zero-shot VLN framework. Within this framework, large models possessing distinct abilities are served as domain experts. Our proposed navigation agent, namely DiscussNav, can actively discuss with these experts to collect essential information before moving at every step. These discussions cover critical navigation subtasks like instruction understanding, environment perception, and completion estimation. Through comprehensive experiments, we demonstrate that discussions with domain experts can effectively facilitate navigation by perceiving instruction-relevant information, correcting inadvertent errors, and sifting through in-consistent movement decisions. The performances on the representative VLN task R2R show that our method surpasses the leading zero-shot VLN model by a large margin on all metrics. Additionally, real-robot experiments display the obvious advantages of our method over single-round self-thinking. Our project web can be seen at the https://sites.google.com/view/discussnav.

We present a novel system which blends multiple distinct sensing modalities in audio-visual surveys to assist marine biologists in collecting datasets for understanding the ecological relationship of fish and other organisms with their habitats on and around coral reefs. Our system, designed for the CUREE AUV, uses four hydrophones to determine the bearing to biological sound sources through beamforming. These observations are merged in a Bayesian Occupancy Grid to produce a 2D map of the acoustic activity of a coral reef. Simultaneously, the AUV uses unsupervised topic modeling to identify different benthic habitats. Combining these maps allows us to determine the level of acoustic activity within each habitat. We demonstrated the system in field trials on reefs in the U.S. Virgin Islands, where it was able to autonomously discover the favored habitats of snapping shrimp (genus Alpheus).

Soft robotic manipulators with many degrees of freedom can carry out complex tasks safely around humans. However, manufacturing of soft robotic hands with several degrees of freedom requires a complex multi-step manual process, which significantly increases their cost. We present a design of a multi-material 15 DoF robotic hand with five fingers including an opposable thumb. Our design has 15 pneumatic actuators based on a series of hollow chambers that are driven by an external pressure system. The thumb utilizes rigid joints and the palm features internal rigid structure and soft skin. The design can be directly 3D printed using a multi-material additive manufacturing process without any assembly process and therefore our hand can be manufactured for less than 300 dollars. We test the hand in conjunction with a low-cost vision-based teleoperation system on different tasks.

Insects have long been recognized for their ability to navigate and return home using visual cues from their nest’s environment. However, the precise mechanism underlying this remarkable homing skill remains a subject of ongoing investigation. Drawing inspiration from the learning flights of honey bees and wasps, we propose a robot navigation method that directly learns the home vector direction from visual percepts during a learning flight in the vicinity of the nest. After learning, the robot will travel away from the nest, come back by means of odometry, and eliminate the resultant drift by inferring the home vector orientation from the currently experienced view. Using a compact convolutional neural network, we demonstrate successful learning in both simulated and real forest environments, as well as successful homing control of a simulated quadrotor. The average errors of the inferred home vectors in general stay well below the 90° required for successful homing, and below 24° if all images contain sufficient texture and illumination. Moreover, we show that the trajectory followed during the initial learning flight has a pronounced impact on the network’s performance. A higher density of sample points in proximity to the nest results in a more consistent return. Code and data are available at https://mavlab.tudelft.nl/learning_to_home.

The ability to plan and control robotic in-hand manipulation is challenged by several issues, including the required amount of prior knowledge of the system and the sophisticated physics that varies across different robot hands or even grasp instances. One of the most direct models of in-hand manipulation is the inverse Jacobian, which can directly map from the desired in-hand object motions to the required hand actuator controls. However, acquiring such inverse Jacobians without complex hand-object system models is typically infeasible. We present a method for controlling in-hand manipulation using inverse Jacobians that are self-identified by a particle filter-based estimation scheme that leverages the ability of underactuated hands to maintain a passively stable grasp during self-identification movements. This method requires no a priori knowledge of the specific hand-object system and learns the system’s inverse Jacobian through small exploratory motions. Our system approximates the underlying inverse Jacobian closely, which can be used to perform manipulation tasks across a range of objects successfully. With extensive experiments on a Yale Model O hand, we show that the proposed system can provide accurate in-hand manipulation of sub-millimeter precision and that the inverse Jacobian-based controller can support real-time manipulation control of up to 900Hz.

Event cameras are recent sensors that measure intensity changes in each pixel asynchronously. It is being used due to lower latency and higher temporal resolution compared to traditional frame-based camera. We propose a method of 3D model-based object tracking directly from events captured by event camera. To enable reliable and accurate tracking of objects, we use a new event representation and predict brightness increment images with motion interpolation. Results of object tracking show the new methods significantly improves tracking duration and robustness, both for perspective and fisheye cameras. Our implementation succeeds in tracking objects when the camera speed is reaching 2 m/s.

The millimeter-wave radar sensor maintains stable performance under adverse environmental conditions, making it a promising solution for all-weather perception tasks, such as outdoor mobile robotics. However, the radar point clouds are relatively sparse and contain massive ghost points, which greatly limits the development of mmWave radar technology. In this paper, we propose a novel point cloud super-resolution approach for 3D mmWave radar data, named Radar-diffusion. Our approach employs the diffusion model defined by mean-reverting stochastic differential equations (SDE). Using our proposed new objective function with supervision from corresponding LiDAR point clouds, our approach efficiently handles radar ghost points and enhances the sparse mmWave radar point clouds to dense LiDAR-like point clouds. We evaluate our approach on two different datasets, and the experimental results show that our method outperforms the state-of-the-art baseline methods in 3D radar super-resolution tasks. Furthermore, we demonstrate that our enhanced radar point cloud is capable of downstream radar point-based registration tasks.

Perceptive deep reinforcement learning (DRL) has lead to many recent breakthroughs for complex AI systems leveraging image-based input data. Applications of these results range from super-human level video game agents to dexterous, physically intelligent robots. However, training these perceptive DRL-enabled systems remains incredibly compute and memory intensive, often requiring huge training datasets and large experience replay buffers. This poses a challenge for the next generation of field robots that will need to be able to learn on the edge in order to adapt to their environments. In this paper, we begin to address this issue through differentially encoded observation spaces. By reinterpreting stored imagebased observations as a video, we leverage lossless differential video encoding schemes to compress the replay buffer without impacting training performance. We evaluate our approach with three state-of-the-art DRL algorithms and find that differential image encoding reduces the memory footprint by as much as 14.2× and 16.7× across tasks from the Atari 2600 benchmark and the DeepMind Control Suite (DMC) respectively. These savings also enable large-scale perceptive DRL that previously required paging between flash and RAM to be run entirely in RAM, improving the latency of DMC tasks by as much as 32%.

Control techniques like MPC can realize contact-rich manipulation which exploits dynamic information, maintaining friction limits and safety constraints. However, contact geometry and dynamics are required to be known. This information is often extracted from CAD, limiting scalability and the ability to handle tasks with varying geometry. To reduce the need for a priori models, we propose a framework for estimating contact models online based on torque and position measurements. To do this, compliant contact models are used, connected in parallel to model multi-point contact and constraints such as a hinge. They are parameterized to be differentiable with respect to all of their parameters (rest position, stiffness, contact location), allowing the coupled robot/environment dynamics to be linearized or efficiently used in gradient-based optimization. These models are then applied for: offline gradient-based parameter fitting, online estimation via an extended Kalman filter, and online gradient-based MPC. The proposed approach is validated on two robots, showing the efficacy of sensorless contact estimation and the effects of online estimation on MPC performance. Video results can be seen at https://youtu.be/CuCTcmn3H-o.

This paper introduces a differentiable representation for the optimization of boustrophedon path plans in convex polygons, explores an additional parameter of these path plans that can be optimized, discusses the properties of this representation that can be leveraged during the optimization process and shows that the previously published attempt at optimization of these path plans was too coarse to be practically useful. Experiments were conducted to show that this differentiable representation can reproduce scores from traditional discrete representations of boustrophedon path plans with high fidelity. Finally, optimization via gradient descent was attempted but found to fail because the search space is far more non-convex than was previously considered in the literature. The wide range of applications for boustrophedon path plans means that this work has the potential to improve path planning efficiency in numerous areas of robotics, including mapping and search tasks using uncrewed aerial systems, environmental sampling tasks using uncrewed marine vehicles, and agricultural tasks using ground vehicles, among numerous others applications.

In this work, we present DiPPeR, a novel and fast 2D path planning framework for quadrupedal locomotion, leveraging diffusion-driven techniques. Our contributions include a scalable dataset generator for map images and corresponding trajectories, an image-conditioned diffusion planner for mobile robots, and a training/inference pipeline employing CNNs. We validate our approach in several mazes, as well as in real-world deployment scenarios on Boston Dynamic’s Spot and Unitree’s Go1 robots. DiPPeR performs on average 23 times faster for trajectory generation against both search based and data driven path planning algorithms with an average of 87% consistency in producing feasible paths of various length in maps of variable size, and obstacle structure. Website: https://rpl-cs-ucl.github.io/DiPPeR/

Recent advances of locomotion controllers utilizing deep reinforcement learning (RL) have yielded impressive results in terms of achieving rapid and robust locomotion across challenging terrain, such as rugged rocks, non-rigid ground, and slippery surfaces. However, while these controllers primarily address challenges underneath the robot, relatively little research has investigated legged mobility through confined 3D spaces, such as narrow tunnels or irregular voids, which impose all-around constraints. The cyclic gait patterns resulted from existing RL-based methods to learn parameterized locomotion skills characterized by motion parameters, such as velocity and body height, may not be adequate to navigate robots through challenging confined 3D spaces, requiring both agile 3D obstacle avoidance and robust legged locomotion. Instead, we propose to learn locomotion skills end-to-end from goal-oriented navigation in confined 3D spaces. To address the inefficiency of tracking distant navigation goals, we introduce a hierarchical locomotion controller that combines a classical planner tasked with planning waypoints to reach a faraway global goal location, and an RL-based policy trained to follow these waypoints by generating low-level motion commands. This approach allows the policy to explore its own locomotion skills within the entire solution space and facilitates smooth transitions between local goals, enabling long-term navigation towards distant goals. In simulation, our hierarchical approach succeeds at navigating through demanding confined 3D environments, outperforming both pure end-to-end learning approaches and parameterized locomotion skills. We further demonstrate the successful real-world deployment of our simulation-trained controller on a real robot.

Recently, reinforcement learning has led to dexterous manipulation skills of increasing complexity. Nonetheless, learning these skills in simulation still exhibits poor sample-efficiency which stems from the fact these skills are learned from scratch without the benefit of any domain expertise. In this work, we aim to improve the sample efficiency of learning dexterous in-hand manipulation skills using controllers available via domain knowledge. To this end, we design simple sub-skill controllers and demonstrate improved sample efficiency using a framework that guides exploration toward relevant state space by following actions from these controllers. We are the first to demonstrate learning hard-to-explore finger-gaiting in-hand manipulation skills without the use of an exploratory reset distribution.

Deformable linear object (DLO) manipulation is needed in many fields. Previous research on deformable linear object (DLO) manipulation has primarily involved parallel jaw gripper manipulation with fixed grasping positions. However, the potential for dexterous manipulation of DLOs using an anthropomorphic hand is under-explored. We present DexDLO, a model-free framework that learns dexterous dynamic manipulation policies for deformable linear objects with a fixed-base dexterous hand in an end-to-end way. By abstracting several common DLO manipulation tasks into goal-conditioned tasks, DexDLO can perform tasks such as DLO grabbing, DLO pulling, DLO end-tip position controlling, etc. Using the Mujoco physics simulator, we demonstrate that our framework can efficiently and effectively learn five different DLO manipulation tasks with the same framework parameters. We further provide a thorough analysis of learned policies, reward functions, and reduced observations for a comprehensive understanding of the framework.