ICRA 2023 Papers

This paper proposes a method for learning continuous control policies for exploration and active landmark localization. We consider a mobile robot detecting landmarks within a limited sensing range, and tackle the problem of learning a control policy that maximizes the mutual information between the landmark states and the sensor observations. We employ a Kalman filter to convert the partially observable problem in the landmark states to a Markov decision process (MDP), a differentiable field of view to shape the reward function, and an attention-based neural network to represent the control policy. The approach is combined with active volumetric mapping to promote environment exploration in addition to landmark localization. The performance is demonstrated in several simulated landmark localization tasks in comparison with benchmark methods.

This paper presents a new technique for learning category-level manipulation from raw RGB-D videos of task demonstrations, with no manual labels or annotations. Category-level learning aims to acquire skills that can be generalized to new objects, with geometries and textures that are different from the ones of the objects used in the demonstrations. We address this problem by first viewing both grasping and manipulation as special cases of tool use, where a tool object is moved to a sequence of key-poses defined in a frame of reference of a target object. Tool and target objects, along with their key-poses, are predicted using a dynamic graph convolutional neural network that takes as input an automatically segmented depth and color image of the entire scene. Empirical results on object manipulation tasks with a real robotic arm show that the proposed network can efficiently learn from real visual demonstrations to perform the tasks on novel objects within the same category, and outperforms alternative approaches.

We present a novel approach for efficient and reliable goal-directed long-horizon navigation for a multi-robot team in a structured, unknown environment by predicting statistics of unknown space. Building on recent work in learning-augmented model based planning under uncertainty, we introduce a high-level state and action abstraction that lets us approximate the challenging Dec-POMDP into a tractable stochastic MDP. Our Multi-Robot Learning over Subgoals Planner (MR-LSP) guides agents towards coordinated exploration of regions more likely to reach the unseen goal. We demonstrate improvement in cost against other multi-robot strategies; in simulated office-like environments, we show that our approach saves 13.29% (2 robot) and 4.6% (3 robot) average cost versus standard non-learned optimistic planning and a learning-informed baseline.

Adaptive falling and recovery skills greatly extend the applicability of robot deployments. In the case of legged mobile manipulators, the robot arm could adaptively stop the fall and assist the recovery. Prior works on falling and recovery strategies for legged mobile manipulators usually rely on assumptions such as inelastic collisions and falling in defined directions to enable real-time computation. This paper presents a learning-based approach to reducing fall damage and recovery. An asymmetric actor-critic training structure is used to train a time-invariant policy with time-varying reward functions. In simulated experiments, the policy recovers from 98.9% of initial falling configurations. It reduces base contact impulse, peak joint internal forces, and base acceleration during the fall compared to the baseline methods. The trained control policy is deployed and extensively tested on the ALMA robot hardware. A video summarizing the proposed method and the hardware tests is available at https://youtu.be/avwg2HqGi8s

Agile flights of autonomous quadrotors in clut-tered environments require constrained motion planning and control subject to translational and rotational dynamics. Tra-ditional model-based methods typically demand complicated design and heavy computation. In this paper, we develop a novel deep reinforcement learning-based method that tackles the challenging task of flying through a dynamic narrow gate. We design a model predictive controller with its adaptive tracking references parameterized by a deep neural network (DNN). These references include the traversal time and the quadrotor SE(3) traversal pose that encourage the robot to fly through the gate with maximum safety margins from various initial conditions. To cope with the difficulty of training in highly dynamic environments, we develop a reinforce-imitate learning framework to train the DNN efficiently that generalizes well to diverse settings. Furthermore, we propose a binary search algorithm that allows online adaption of the SE(3) references to dynamic gates in real-time. Finally, through extensive high-fidelity simulations, we show that our approach is adaptive to different gate trajectories, velocities, and orientations.

Interactions with articulated objects are a challenging but important task for mobile robots. To tackle this challenge, we propose a novel closed-loop control pipeline, which integrates manipulation priors from affordance estimation with sampling-based whole-body control. We introduce the concept of agent-aware affordances which fully reflect the agent's capabilities and embodiment and we show that they outperform their state-of-the-art counterparts which are only conditioned on the end-effector geometry. Additionally, closed-loop affordance inference is found to allow the agent to divide a task into multiple non-continuous motions and recover from failure and unexpected states. Finally, the pipeline is able to perform long-horizon mobile manipulation tasks, i.e. opening and closing an oven, in the real world with high success rates (opening: 71%, closing: 72%).

We present Learned Risk Metric Maps (LRMM) for real-time estimation of coherent risk metrics of high-dimensional dynamical systems operating in unstructured, partially observed environments. LRMM models are simple to design and train-requiring only procedural generation of obstacle sets, state and control sampling, and supervised training of a function approximator-which makes them broadly applicable to arbitrary system dynamics and obstacle sets. In a parallel autonomy setting, we demonstrate the model's ability to rapidly infer collision probabilities of a fast-moving car-like robot driving recklessly in an obstructed environment; allowing the LRMM agent to intervene, take control of the vehicle, and avoid collisions. In this time-critical scenario, we show that LRMMs can evaluate risk metrics 20-100x times faster than alternative safety algorithms based on control barrier functions (CBFs) and Hamilton-Jacobi reachability (HJ-reach), leading to 5–15 % fewer obstacle collisions by the LRMM agent than CBFs and HJ-reach. This performance improvement comes in spite of the fact that the LRMM model only has access to local/partial observation of obstacles, whereas the CBF and HJ-reach agents are granted privileged/global information. We also show that our model can be equally well trained on a 12-dimensional quadrotor system operating in an obstructed indoor environment. The LRMM codebase is provided at https://github.com/mit-drl/pyrmm.

Central Pattern generators (CPG) are a biologically inspired, decentralized control architecture that enables model-free, but yet adaptively stable and computational lightweight locomotion capabilities on complex robots. Nevertheless, no unified design guidelines for closed-loop CPG controllers are available in the literature. Therefore, we propose a task-distributed, end-to-end trainable, closed-loop CPG control policy by generalizing and extending Tegotae control. The Tegotae approach modulates CPG activity by quantifying the discrepancy between internal belief states and environmental reactions. Spontaneous and adaptive gait formation towards situationally efficient locomotion patterns are intrinsic properties of Tegotae control. The Tegotae control policy is trained and benchmarked in simulation on a 1D hopping robot. We found that our approach can learn efficient and adaptive locomotion on minimal feedback information, while out-performing unstructured, classic reinforcement learning policies of equal complexity. To the best of our knowledge, this is the first study to fully generalize the Tegotae approach and construct unimpeded, end-to-end trainable Tegotae control policies.

In this paper, we propose using online public maps, e.g., OpenStreetMap (OSM), for large-scale radar-based localization without needing a prior sensing map. This can potentially extend the localization system to anywhere worldwide without building, saving, or maintaining a sensing map, as long as an online public map covers the operating area. Existing methods using OSM only use route network or semantics information. These two sources of information are not combined in the previous works, while our proposed system fuses them to improve localization accuracy. Our experiments, on three open datasets collected from three different continents, show that the proposed system outperforms the state-of-the-art localization methods, reducing up to 50% of position errors. We release an open-source implementation for the community.

Scene completion refers to obtaining dense scene representation from an incomplete perception of complex 3D scenes. This helps robots detect multi-scale obstacles and analyse object occlusions in scenarios such as autonomous driving. Recent advances show that implicit representation learning can be leveraged for continuous scene completion and achieved through physical constraints like Eikonal equations. However, former Eikonal completion methods only demonstrate results on watertight meshes at a scale of tens of meshes. None of them are successfully done for non-watertight LiDAR point clouds of open large scenes at a scale of thousands of scenes. In this paper, we propose a novel Eikonal formulation that conditions the implicit representation on localized shape priors which function as dense boundary value constraints, and demonstrate it works on SemanticKITTI and SemanticPOSS. It can also be extended to semantic Eikonal scene completion with only small modifications to the network architecture. With extensive quantitative and qualitative results, we demonstrate the benefits and drawbacks of existing Eikonal methods, which naturally leads to the new locally conditioned formulation. Notably, we improve IoU from 31.7% to 51.2% on SemanticKITTI and from 40.5% to 48.7% on SemanticPOSS. We extensively ablate our methods and demonstrate that the proposed formulation is robust to a wide spectrum of implementation hyper-parameters. Codes and models are publicly available at https://github.com/AIR-DISCOVER/LODE

Correspondence selection, a crucial step in many computer vision tasks, aims to distinguish between inliers and outliers from putative correspondences. The coherence of correspondences is often used for predicting inlier probability, but it is difficult for neural networks to extract coherence contexts based only on quadruple coordinates. To overcome this difficulty, we propose enhancing the preliminary features using local and global handcrafted coherent characteristics before model learning, which strengthens the discrimination of each correspondence and guides the model to prune obvious outliers. Furthermore, to fully utilize local information, neighbors are searched in coordinate space as well as feature space. These two kinds of neighbors provide complementary and plentiful contexts for inlier probability prediction. Finally, a novel neighbor representation and a fusion architecture are proposed to retain detailed features. Experiments demonstrate that our method achieves state-of-the-art performance on relative camera pose estimation and correspondence selection metrics on the outdoor YFCC100M [1] and the indoor SUN3D [2] datasets.

Sampling-based algorithms solve the path planning problem by generating random samples in the searchspace and incrementally growing a connectivity graph or a tree. Conventionally, the sampling strategy used in these algorithms is biased towards exploration to acquire information about the search-space. In contrast, this work proposes an optimization-based procedure that generates new samples so as to improve the cost-to-come value of vertices in a given neighborhood. The application of the proposed algorithm adds an exploitativebias to sampling and results in a faster convergence to the optimal solution compared to other state-of-the-art sampling techniques. This is demonstrated using benchmarking experiments performed for 7 DOF Panda and 14 DOF Baxter robots.

This paper presents LEMURS, an algorithm for learning scalable multi-robot control policies from cooperative task demonstrations. We propose a port-Hamiltonian description of the multi-robot system to exploit universal physical constraints in interconnected systems and achieve closed-loop stability. We represent a multi-robot control policy using an architecture that combines self-attention mechanisms and neural ordinary differential equations. The former handles time-varying communication in the robot team, while the latter respects the continuous-time robot dynamics. Our representation is distributed by construction, enabling the learned control policies to be deployed in robot teams of different sizes. We demonstrate that LEMURS can learn interactions and cooperative behaviors from demonstrations of multi-agent navigation and flocking tasks.

State estimation is an important aspect in many robotics applications. In this work, we consider the task of obtaining accurate state estimates for robotic systems by enhancing the dynamics model used in state estimation algorithms. Existing frameworks such as moving horizon estimation (MHE) and the unscented Kalman filter (UKF) provide the flexibility to incorporate nonlinear dynamics and measurement models. However, this implies that the dynamics model within these algorithms has to be sufficiently accurate in order to warrant the accuracy of the state estimates. To enhance the dynamics models and improve the estimation accuracy, we utilize a deep learning framework known as knowledge-based neural ordinary differential equations (KNODEs). The KNODE framework embeds prior knowledge into the training procedure and synthesizes an accurate hybrid model by fusing a prior first-principles model with a neural ordinary differential equation (NODE) model. In our proposed LEARNEST framework, we integrate the data-driven model into two novel model-based state estimation algorithms, which are denoted as KNODE-MHE and KNODE-UKF. These two algorithms are compared against their conventional counterparts across a number of robotic applications; state estimation for a cartpole system using partial measurements, localization for a ground robot, as well as state estimation for a quadrotor. Through simulations and tests using real-world experimental data, we demonstrate the versatility and efficacy of the proposed learning-enhanced state estimation framework.

Natural language is one of the most intuitive ways to express human intent. However, translating instructions and commands towards robotic motion generation and deployment in the real world is far from being an easy task. The challenge of combining a robot's inherent low-level geometric and kinodynamic constraints with a human's high-level semantic instructions traditionally is solved using task-specific solutions with little generalizability between hardware platforms, often with the use of static sets of target actions and commands. This work instead proposes a flexible language-based framework that allows a user to modify generic robotic trajectories. Our method leverages pre-trained language models (BERT and CLIP) to encode the user's intent and target objects directly from a free-form text input and scene images, fuses geometrical features generated by a transformer encoder network, and finally outputs trajectories using a transformer decoder, without the need of priors related to the task or robot information. We significantly extend our own previous work presented in [1] by expanding the trajectory parametrization space to 3D and velocity as opposed to just XY movements. In addition, we now train the model to use actual images of the objects in the scene for context (as opposed to textual descriptions), and we evaluate the system in a diverse set of scenarios beyond manipulation, such as aerial and legged robots. Our simulated and real-life experiments demonstrate that our transformer model can successfully follow human intent, modifying the shape and speed of trajectories within multiple environments. Codebase avail-able at: https://github.com/arthurfenderbucker/LaTTe-Language-Trajectory-TransformEr.git.

Neural Radiance Fields (NeRFs) have made great success in representing complex 3D scenes with high-resolution details and efficient memory. Nevertheless, current NeRF - based pose estimators have no initial pose prediction and are prone to local optima during optimization. In this paper, we present LATITUDE: Global Localization with Truncated Dynamic Low-pass Filter, which introduces a two-stage localization mechanism in city-scale NeRF. In place recognition stage, we train a regressor through images generated from trained NeRFs, which provides an initial value for global localization. In pose optimization stage, we minimize the residual between the observed image and rendered image by directly optimizing the pose on the tangent plane. To avoid falling into local optimum, we introduce a Truncated Dynamic Low-pass Filter (TDLF) for coarse-to-fine pose registration. We evaluate our method on both synthetic and real-world data and show its potential applications for high-precision navigation in large-scale city scenes. Codes and dataset will be publicly available at https://github.com/jike5/LATITUDE.

Semantic grids can be useful representations of the scene around an autonomous system. By having information about the layout of the space around itself, a robot can leverage this type of representation for crucial tasks such as navigation or tracking. By fusing information from multiple sensors, robustness can be increased and the computational load for the task can be lowered, achieving real time performance. Our multi-scale LiDAR-Aided Perspective Transform network uses information available in point clouds to guide the projection of image features to a top-view representation, resulting in a relative improvement in the state of the art for semantic grid generation for human (+8.67%) and movable object (+49.07%) classes in the nuScenes dataset, as well as achieving results close to the state of the art for the vehicle, drivable area and walkway classes, while performing inference at 25 FPS.

As neuromorphic technology is maturing, its application to robotics and autonomous vehicle systems has become an area of active research. In particular, event cameras have emerged as a compelling alternative to frame-based cameras in low-power and latency-demanding applications. To enable event cameras to operate alongside staple sensors like lidar in perception tasks, we propose a direct, temporally-decoupled extrinsic calibration method between event cameras and lidars. The high dynamic range, high temporal resolution, and low-latency operation of event cameras are exploited to directly register lidar laser returns, allowing information-based correlation methods to optimize for the 6- DoF extrinsic calibration between the two sensors. This paper presents the first direct calibration method between event cameras and lidars, removing dependencies on frame-based camera intermediaries and/or highly-accurate hand measurements. Code: https://github.com/kev-in-ta/12e

This study presents a framework, L-C*, for resilient and lightweight direct visual localization, employing a loosely coupled fusion of visual and inertial data. Unlike indirect methods, direct visual localization facilitates accurate pose estimation on general color three-dimensional maps that are not tailored for visual localization. However, it suffers from temporal localization failures and high computational costs for real-time applications. For long-term and real-time visual localization, we developed an L-C* that incorporates direct visual localization C* in a visual-inertial loose coupling. By capturing ego-motion via visual-inertial odometry to interpolate global pose estimates, the framework allows for a significant reduction in the frequency of demanding global localization, thereby facilitating lightweight but reliable visual localization. In addition, forming a closed loop that feeds the latest pose estimate to the visual localization component as an initial guess for the next pose inference renders the system highly robust. A quantitative evaluation of a simulation dataset demonstrated the accuracy and efficiency of the proposed framework. Experiments using smartphone sensors also demonstrated the robustness and resiliency of L-C* in real-world situations.