RSS 2024 论文列表

Recent strides in model predictive control (MPC) underscore a dependence on numerical advancements to efficiently and accurately solve large-scale problems. Given the substantial number of variables characterizing typical whole-body optimal control problems —often numbering in the thousands— exploiting the sparse structure of the numerical problem becomes crucial to meet computational demands, typically in the range of a few milliseconds. A fundamental building block for computing Newton or Sequential Quadratic Programming steps in direct optimal control methods involves addressing the linear-quadratic regulator (LQR) problem. This paper concentrates on equality-constrained problems featuring implicit system dynamics and dual regularization, a characteristic found in advanced interior-point or augmented-Lagrangian solvers. Here, we introduce a parallel algorithm designed for solving an LQR problem with dual regularization. Leveraging a rewriting of the LQR recursion through block elimination, we first enhanced the efficiency of the serial algorithm, then subsequently generalized to handle parametric LQR. This extension enables us to split decision variables and to solve several subproblems concurrently. Our algorithm is implemented (and will be released in open source) in a nonlinear constrained implicit optimal control solver. It showcases improved performance over previous serial formulations and we validate its efficacy by deploying it in the model predictive control of a real quadruped robot.

In this paper, we seek to learn a robot policy guaranteed to satisfy state constraints. To encourage constraint satisfaction, existing RL algorithms typically rely on Constrained Markov Decision Processes and discourage constraint violations through reward shaping. However, such soft constraints cannot offer safety guarantees. To address this gap, we propose POLICEd RL, a novel RL algorithm explicitly designed to enforce affine hard constraints in closed-loop with a black-box environment. Our key insight is to make the learned policy be affine around the unsafe set and to use this affine region as a repulsive buffer to prevent trajectories from violating the constraint. We prove that such policies exist and guarantee constraint satisfaction. Our proposed framework is applicable to both systems with continuous and discrete state and action spaces and is agnostic to the choice of the RL training algorithm. Our results demonstrate the capacity of POLICEd RL to enforce hard constraints in robotic tasks while significantly outperforming existing methods. Code available at https://iconlab.negarmehr.com/POLICEd-RL/

Gaussian Process (GP) models are widely used for Robotic Information Gathering (RIG) in exploring unknown environments due to their ability to model complex phenomena with non-parametric flexibility and accurately quantify prediction uncertainty. Previous work has developed informative planners and adaptive GP models to enhance the data efficiency of RIG by improving the robot’s sampling strategy to focus on informative regions in non-stationary environments. However, computational efficiency becomes a bottleneck when using GP models in large-scale environments with limited computational resources. We propose a framework – Probabilistic Online Attentive Mapping (POAM) – that leverages the modeling strengths of the non-stationary Attentive Kernel while achieving constant-time computational complexity for online decision-making. POAM guides the optimization process via variational Expectation Maximization, providing constant-time update rules for inducing inputs, variational parameters, and hyperparameters. Extensive experiments in active bathymetric mapping tasks demonstrate that POAM significantly improves computational efficiency, model accuracy, and uncertainty quantification capability compared to existing online sparse GP models.

A generalised solver for the manipulator non-revisiting coverage path planning (NCPP) problem is proposed in this paper. Nonlinear manipulator kinematics and the imposition of task-specific constraints dictate that applying conventional coverage path planning (CPP) solutions based on 2D template matching or cellular decomposition schemes on the target surface invariably results in truncated end-effector motions. Likewise, coverage paths designed directly in joint-space cannot ensure re-occurrences will not arise. More recent SOTA works have proposed finite-step optimal NCPP solutions where singularities are however expressly disregarded. Directly incorporating singular configurations violates the local bijectivity and finite-to-one property in the kinematic mapping, and cannot be properly modelled within existing schemes. This work leverages "valid'' singularities, those that exhibit sufficient manoeuvrability in suitable dimensions to allow continuation of the tracking motion, thus further reducing the number of posture reconfigurations. The scheme assumes a generic representation of surfaces as discrete meshes, symbolising a null probability to locate the points corresponding to valid but singular inverse kinematic configurations, and constructs a practical method to traverse a singularity without explicitly calculating it. Simulated and realistic experiments are carried out where the suitability of the scheme to reduce posture reconfigurations and achieve continuous coverage motions are compared with existing methods. Three scenarios have been examined whereby the planner is able to fulfil motions without discontinuities in all instances.

Learning a single universal policy that can perform a diverse set of manipulation tasks is a promising new direction in robotics. However, existing techniques are limited to learning policies that can only perform tasks that are encountered during training, and require a large number of demonstrations to learn new tasks. Humans, on the other hand, often can learn a new task from a single unannotated demonstration. In this work, we propose the Invariance-Matching One-shot Policy Learning (IMOP) algorithm. In contrast to the standard practice of learning the end-effector’s pose directly, IMOP first learns invariant regions of the state space for a given task, and then computes the end-effector’s pose through matching the invariant regions between demonstrations and test scenes. Trained on the 18 RLBench tasks, IMOP achieves a success rate that outperforms the state- of-the-art consistently, by 4.5% on average over the 18 tasks. More importantly, IMOP can learn a novel task from a single unannotated demonstration, and without any fine-tuning, and achieves an average success rate improvement of 11.5% over the state-of-the-art on 22 novel tasks selected across nine categories. IMOP can also generalize to new shapes and learn to manipulate objects that are different from those in the demonstration. Further, IMOP can perform one-shot sim-to-real transfer using a single real-robot demonstration.

Imitation learning is a powerful machine learning algorithm for a robot to acquire manipulation skills. Nevertheless, many real-world manipulation tasks involve precise and dexterous robot-object interactions, which make it difficult for humans to collect high-quality expert demonstrations. As a result, a robot has to learn skills from suboptimal demonstrations and unstructured interactions, which remains a key challenge. Existing works typically use offline deep reinforcement learning~(RL) to solve this challenge, but in practice these algorithms are unstable and fragile due to the deadly triad issue. To overcome this problem, we propose GSR, a simple yet effective algorithm that learns from suboptimal demonstrations through Graph Search and Retrieval. We first use pretrained representation to organize the interaction experience into a graph and perform a graph search to calculate the values of different behaviors. Then, we apply a retrieval-based procedure to identify the best behavior (actions) on each state and use behavior cloning to learn that behavior. We evaluate our method in both simulation and real-world robotic manipulation tasks with complex visual inputs, covering various precise and dexterous manipulation skills with objects of different physical properties. GSR can achieve a 10% to 30% higher success rate and over 30% higher proficiency compared to baselines.

Physical reasoning is important for effective robot manipulation. Recent work has investigated both vision and language modalities for physical reasoning; vision can reveal information about objects in the environment and language serves as an abstraction and communication medium for additional context. Although these works have demonstrated success on a variety of physical reasoning tasks, they are limited to physical properties that can be inferred from visual or language inputs. In this work, we investigate combining tactile perception with language, which enables embodied systems to obtain physical properties through interaction and apply commonsense reasoning. We contribute a new dataset PhysiCLeAR, which comprises both physical/property reasoning tasks and annotated tactile videos obtained using a GelSight tactile sensor. We then introduce Octopi, a system that leverages both tactile representation learning and large vision-language models to predict and reason about tactile inputs with minimal language fine-tuning. Our evaluations on PhysiCLeAR show that Octopi is able to effectively use intermediate physical property predictions to improve its performance on various tactile-related tasks. PhysiCLeAR and Octopi are available at https://github.com/clear-nus/octopi.

Large policies pretrained on diverse robot datasets have the potential to transform robotic learning: instead of training new policies from scratch, such generalist robot policies may be finetuned with only a little in-domain data, yet generalize broadly. However, to be widely applicable across a range of robotic learning scenarios, environments, and tasks, such policies need to handle diverse sensors and action spaces, accommodate a variety of commonly used robotic platforms, and finetune readily and efficiently to new domains. In this work, we aim to lay the groundwork for developing open-source, widely applicable, generalist policies for robotic manipulation. As a first step, we introduce Octo, a large transformer-based policy trained on 800k trajectories from the Open X-Embodiment dataset, the largest robot manipulation dataset to date. It can be instructed via language commands or goal images and can be effectively finetuned to robot setups with new sensory inputs and action spaces within a few hours on standard consumer GPUs. In experiments across 9 robotic platforms, we demonstrate that Octo serves as a versatile policy initialization that can be effectively finetuned to new observation and action spaces. We also perform detailed ablations of design decisions for the Octo model, from architecture to training data, to guide future research on building generalist robot models.

The main challenge in learning image-conditioned robotic policies is acquiring a visual representation conducive to low-level control. Due to the high dimensionality of the image space, learning a good visual representation requires a considerable amount of visual data. However, when learning in the real world, data is expensive. Sim2Real is a promising paradigm for overcoming data scarcity in the real-world target domain by using a simulator to collect large amounts of cheap data closely related to the target task. However, it is difficult to transfer an image-conditioned policy from sim to real when the domains are very visually dissimilar. To bridge the sim2real visual gap, we propose using natural language descriptions of images as a unifying signal across domains that captures the underlying task-relevant semantics. Our key insight is that if two image observations from different domains are labeled with similar language, the policy should predict similar action distributions for both images. We demonstrate that training the image encoder to predict the language description or the distance between descriptions of a sim or real image serves as a useful, data-efficient pretraining step that helps learn a domain-invariant image representation. We can then use this image encoder as the backbone of an IL policy trained simultaneously on a large amount of simulated and a handful of real demonstrations. Our approach outperforms widely used prior sim2real methods and strong vision-language pretraining baselines like CLIP and R3M by 25 to 40\%. See additional videos and materials at https://robin-lab.cs.utexas.edu/lang4sim2real/.

Vision-and-language navigation (VLN) stands as a key research problem of Embodied AI, aiming at enabling agents to navigate in unseen environments following linguistic instructions. In this field, generalization is a long-standing challenge, either to out-of-distribution scenes or from Sim to Real. In this paper, we propose NaVid, a video-based large vision language model (VLM), to mitigate such a generalization gap. NaVid makes the first endeavor to showcase the capability of VLMs to achieve state-of-the-art level navigation performance without any maps, odometers, or depth inputs. Following human instruction, NaVid only requires an on-the-fly video stream from a monocular RGB camera equipped on the robot to output the next-step action. Our formulation mimics how humans navigate and naturally gets rid of the problems introduced by odometer noises, and the Sim2Real gaps from map or depth inputs. Moreover, our video-based approach can effectively encode the historical observations of robots as spatio-temporal contexts for decision making and instruction following. We train NaVid with 510k navigation samples collected from continuous environments, including action-planning and instruction-reasoning samples, along with 763k large-scale web data. Extensive experiments show that NaVid achieves state-of-the-art performance in simulation environments and the real world, demonstrating superior cross-dataset and Sim2Real transfer. We thus believe our proposed VLM approach plans the next step for not only the navigation agents but also this research field.

This work introduces the Multimodal Diffusion Transformer (MDT), a novel diffusion policy framework, that excels at learning versatile behavior from multimodal goal specifications with few language annotations. MDT leverages a diffusion based multimodal transformer backbone and two self-supervised auxiliary objectives to master long-horizon manipulation tasks based on multimodal goals. The vast majority of imitation learning methods only learn from individual goal modalities, e.g. either language or goal images. However, existing large-scale imitation learning datasets are only partially labeled with language annotations, which prohibits current methods from learning language conditioned behavior from these datasets. MDT addresses this challenge by introducing a latent goal-conditioned state representation, that is simultaneously trained on multimodal goal instructions. This state representation aligns image and language based goal embeddings and encodes sufficient information to predict future states. The representation is trained via two self-supervised auxiliary objectives that enhance the performance of the presented transformer backbone. MDT shows exceptional performance on 164 tasks provided by the challenging CALVIN and LIBERO benchmarks, including a LIBERO version that contains less than $2\\%$ language annotations. Further, MDT establishes a new record on the CALVIN manipulation challenge, demonstrating an absolute performance improvement of $15\\%$ over prior state-of-the-art methods, that require large-scale pretraining and contain $10\times$ more learnable parameters. MDT demonstrated its ability to solve long-horizon manipulation from sparsely annotated data in both simulated and real-world environments. Demonstrations and Code are available at https://intuitive-robots.github.io/mdt_policy/.

Numerous classes of robotics motion planning problems involve searching in constrained configuration spaces where the constraints change during different stages of the motion, and these kinds of motion planning problems are named multi-modal problems. The most common method to solve these problems is to represent them as a set of manifolds and search for a trajectory across them. Often, instead of using manifolds alone, foliated manifolds, which are a union of disjoint manifolds, are a better way to model the manipulation problem. However, the complexity of planning in foliated manifolds is significant due to the increased number of manifolds, hard task constraints, and complex environments. To tackle these challenges, we propose an efficient planning framework that leverages a dynamic roadmap structure to learn from accumulated experience acquired during previous planning attempts in similar foliated manifolds. When planning in a new foliated manifold, this experience, captured in configuration distributions and an atlas, which are tangential charts approximating the new manifold with constraints, is effectively utilized to guide motion planning. We demonstrate the framework's performance for manipulation problems with different foliated manifold structures in simulation and real-world scenarios. An open-source will be released soon.

Terrain traversability in unstructured off-road autonomy has traditionally relied on semantic classification, resource-intensive dynamics models, or purely geometry-based methods to predict vehicle-terrain interactions. While inconsequential at low speeds, uneven terrain subjects our full-scale system to safety-critical challenges at operating speeds of 7-10 m/s. This study focuses particularly on uneven terrain such as hills, banks, and ditches. These common high-risk geometries are capable of disabling the vehicle and causing severe passenger injuries if poorly traversed. We introduce a physics-based framework for identifying traversability constraints on terrain dynamics. Using this framework, we derive two fundamental constraints, each with a focus on mitigating rollover and ditch-crossing failures while being fully parallelizable in the sample-based Model Predictive Control (MPC) framework. In addition, we present the design of our planning and control system, which implements our parallelized constraints in MPC and utilizes a low-level controller to meet the demands of our aggressive driving without prior information about the environment and its dynamics. Through real-world experimentation and traversal of hills and ditches, we demonstrate that our approach captures fundamental elements of safe and aggressive autonomy over uneven terrain. Our approach improves upon geometry-based methods by completing comprehensive off-road courses up to 22\% faster while maintaining safe operation.

Quadrotor flight is an extremely challenging problem due to the limited control authority encountered at the limit of handling. Model Predictive Contouring Control (MPCC) has emerged as a promising model-based approach for time optimization problems such as drone racing. However, the standard MPCC formulation used in quadrotor racing introduces the notion of the gates directly in the cost function, creating a multi-objective optimization that continuously trades off between maximizing progress and tracking the path accurately. This paper introduces three key components that enhance the state-of-the-art MPCC approach for drone racing. First and foremost, we provide safety guarantees in the form of a track constraint and terminal set. The track constraint is designed as a spatial constraint which prevents gate collisions while allowing for time optimization only in the cost function. Second, we augment the existing first principles dynamics with a residual term that captures complex aerodynamic effects and thrust forces learned directly from real-world data. Third, we use Trust Region Bayesian Optimization (TuRBO), a state-of-the-art global Bayesian Optimization algorithm, to tune the hyperparameters of the MPCC controller given a sparse reward based on lap time minimization. The proposed approach achieves similar lap times to the best-performing RL policy and outperforms the best model-based controller while satisfying constraints. In both simulation and real world, our approach consistently prevents gate crashes with 100% success rate, while pushing the quadrotor to its physical limits reaching speeds of more than 80km/h.

Open-world generalization requires robotic systems to have a profound understanding of the physical world and the user command to solve diverse and complex tasks. While the recent advancement in vision-language models (VLMs) has offered unprecedented opportunities to solve open-world problems, how to leverage their capabilities to control robots remains a grand challenge. In this paper, we introduce Marking Open-world Keypoint Affordances (MOKA), an approach that employs VLMs to solve robotic manipulation tasks specified by free-form language instructions. Central to our approach is a compact point-based representation of affordance, which bridges the VLM’s predictions on observed images and the robot’s actions in the physical world. By prompting the pre-trained VLM, our approach utilizes the VLM’s commonsense knowledge and concept understanding acquired from broad data sources to predict affordances and generate motions. To facilitate the VLM’s reasoning in zero-shot and few-shot manners, we propose a visual prompting technique that annotates marks on images, converting affordance reasoning into a series of visual question-answering problems that are solvable by the VLM. We further explore methods to enhance performance with robot experiences collected by MOKA through in-context learning and policy distillation. We evaluate and analyze MOKA’s performance on various table-top manipulation tasks including tool use, deformable body manipulation, and object rearrangement.

The ability to reuse collected data and transfer trained policies between robots could alleviate the burden of additional data collection and training. While existing approaches such as pretraining plus finetuning and co-training show promise, they do not generalize to robots unseen in training. Focusing on common robot arms with similar workspaces and 2-jaw grippers, we investigate the feasibility of zero-shot transfer. Through simulation studies on 8 manipulation tasks, we find that state-based Cartesian control policies can successfully zero-shot transfer to a target robot after accounting for forward dynamics. To address robot visual disparities for vision-based policies, we introduce Mirage, which uses “cross-painting”—masking out the unseen target robot and inpainting the seen source robot—during execution in real time so that it appears to the policy as if the trained source robot were performing the task. Mirage applies to both first-person and third-person camera views and policies that take in both states and images as inputs or only images as inputs. Despite its simplicity, our extensive simulation and physical experiments provide strong evidence that Mirage can successfully zero-shot transfer between different robot arms and grippers with only minimal performance degradation on a variety of manipulation tasks such as picking, stacking, and assembly, significantly outperforming a generalist policy.

Recent advances in robot skill learning have unlocked the potential to construct task-agnostic skill libraries, facilitating the seamless sequencing of multiple simple manipulation primitives (aka. skills) to tackle significantly more complex tasks. Nevertheless, determining the optimal sequence for independently learned skills remains an open problem, particularly when the objective is given solely in terms of the final geometric configuration rather than a symbolic goal. To address this challenge, we propose Logic-Skill Programming (LSP), an optimization-based approach that sequences independently learned skills to solve long-horizon tasks. We formulate a first-order extension of a mathematical program to optimize the overall cumulative reward of all skills within a plan, abstracted by the sum of value functions. To solve such programs, we leverage the use of tensor train factorization to construct the value function space, and rely on alternations between symbolic search and skill value optimization to find the appropriate skill skeleton and optimal subgoal sequence. Experimental results indicate that the obtained value functions provide a superior approximation of cumulative rewards compared to state-of-the-art reinforcement learning methods. Furthermore, we validate LSP in three manipulation domains, encompassing both prehensile and non-prehensile primitives. The results demonstrate its capability to identify the optimal solution over the full logic and geometric path. The real-robot experiments showcase the effectiveness of our approach to cope with contact uncertainty and external disturbances in the real world.

For decades, inverse kinematics (IK) was an intense and active research area in robotics. Beyond analytical solutions limited to a restricted range of robotic systems and applications, differential inverse kinematics has emerged as a generic class of methods, able to cope with a wider variety of robots and scenarios, with quadratic programming-based approaches as the main paradigm. In this paper, we propose to revisit differential inverse kinematics from the perspective of augmented Lagrangian methods (AL) and the well-known related alternating direction method of multipliers (ADMM). Notably, by leveraging AL techniques and in the spirit of Featherstone algorithms, we introduce a rigid-body dynamics algorithm that solves equality-constrained IK problems with linear complexity in the number of robot joints and number of constraints. Combined with the ADMM strategy developed in the OSQP solver, we provide a new solution for the same class of problems as QP-based differential IK, yet with linear complexity in problem dimensions. We propose an open-source C++ implementation of this approach, which we validate on a large set of problems including manipulation and humanoid locomotion tasks. Our benchmark measures computation times 2--3 $\times$ shorter than the QP-based state of the art.

Compared with the widely investigated homogeneous multi-robot collaboration, heterogeneous robots with different capabilities can provide a more efficient and flexible collaboration for more complex tasks. In this paper, we consider a more challenging heterogeneous ad hoc teamwork collaboration problem where an ad hoc robot joins an existing heterogeneous team for a shared goal. Specifically, the ad hoc robot collaborates with unknown teammates without prior coordination, and it is expected to generate an appropriate cooperation policy to improve the efficiency of the whole team. To solve this challenging problem, we leverage the remarkable potential of the large language model (LLM) to establish a decentralized heterogeneous ad hoc teamwork collaboration framework that focuses on generating reasonable policy for an ad hoc robot to collaborate with original heterogeneous teammates. A training-free hierarchical dynamic planner is developed using the LLM together with the newly proposed Interactive Reflection of Thoughts (IRoT) method for the ad hoc agent to adapt to different teams. We also build a benchmark testing dataset to evaluate the proposed framework in the heterogeneous ad hoc multi-agent tidying-up task. Extensive comparison and ablation experiments are conducted in the benchmark to demonstrate the effectiveness of the proposed framework. We have also employed the proposed framework in physical robots in a real-world scenario. The experimental videos can be found at https://youtu.be/wHYP5T2WIp0.