Weiming Zhi

Tin Lai, Weiming Zhi, Tucker Hermans et al. · nvidia

We present the Learning for KinoDynamic Tree Expansion (L4KDE) method for kinodynamic planning. Tree-based planning approaches, such as rapidly exploring random tree (RRT), are the dominant approach to finding globally optimal plans in continuous state-space motion planning. Central to these approaches is tree-expansion, the procedure in which new nodes are added into an ever-expanding tree. We study the kinodynamic variants of tree-based planning, where we have known system dynamics and kinematic constraints. In the interest of quickly selecting nodes to connect newly sampled coordinates, existing methods typically cannot optimise to find nodes that have low cost to transition to sampled coordinates. Instead, they use metrics like Euclidean distance between coordinates as a heuristic for selecting candidate nodes to connect to the search tree. We propose L4KDE to address this issue. L4KDE uses a neural network to predict transition costs between queried states, which can be efficiently computed in batch, providing much higher quality estimates of transition cost compared to commonly used heuristics while maintaining almost-surely asymptotic optimality guarantee. We empirically demonstrate the significant performance improvement provided by L4KDE on a variety of challenging system dynamics, with the ability to generalise across different instances of the same model class, and in conjunction with a suite of modern tree-based motion planners.

Weiming Zhi, Tianyi Zhang, Matthew Johnson-Roberson

Learning for Demonstration (LfD) enables robots to acquire new skills by imitating expert demonstrations, allowing users to communicate their instructions in an intuitive manner. Recent progress in LfD often relies on kinesthetic teaching or teleoperation as the medium for users to specify the demonstrations. Kinesthetic teaching requires physical handling of the robot, while teleoperation demands proficiency with additional hardware. This paper introduces an alternative paradigm for LfD called Diagrammatic Teaching. Diagrammatic Teaching aims to teach robots novel skills by prompting the user to sketch out demonstration trajectories on 2D images of the scene, these are then synthesised as a generative model of motion trajectories in 3D task space. Additionally, we present the Ray-tracing Probabilistic Trajectory Learning (RPTL) framework for Diagrammatic Teaching. RPTL extracts time-varying probability densities from the 2D sketches, applies ray-tracing to find corresponding regions in 3D Cartesian space, and fits a probabilistic model of motion trajectories to these regions. New motion trajectories, which mimic those sketched by the user, can then be generated from the probabilistic model. We empirically validate our framework both in simulation and on real robots, which include a fixed-base manipulator and a quadruped-mounted manipulator.

Weiming Zhi, Tianyi Zhang, Matthew Johnson-Roberson

Diagrammatic Teaching is a paradigm for robots to acquire novel skills, whereby the user provides 2D sketches over images of the scene to shape the robot's motion. In this work, we tackle the problem of teaching a robot to approach a surface and then follow cyclic motion on it, where the cycle of the motion can be arbitrarily specified by a single user-provided sketch over an image from the robot's camera. Accordingly, we contribute the Stable Diffeomorphic Diagrammatic Teaching (SDDT) framework. SDDT models the robot's motion as an Orbitally Asymptotically Stable (O.A.S.) dynamical system that learns to stablize based on a single diagrammatic sketch provided by the user. This is achieved by applying a \emph{diffeomorphism}, i.e. a differentiable and invertible function, to morph a known O.A.S. system. The parameterised diffeomorphism is then optimised with respect to the Hausdorff distance between the limit cycle of our modelled system and the sketch, to produce the desired robot motion. We provide novel theoretical insight into the behaviour of the optimised system and also empirically evaluate SDDT, both in simulation and on a quadruped with a mounted 6-DOF manipulator. Results show that we can diagrammatically teach complex cyclic motion patterns with a high degree of accuracy.

Tianyi Zhang, Weiming Zhi, Kaining Huang et al.

Water caustics are commonly observed in seafloor imaging data from shallow-water areas. Traditional methods that remove caustic patterns from images often rely on 2D filtering or pre-training on an annotated dataset, hindering the performance when generalizing to real-world seafloor data with 3D structures. In this paper, we present a novel method Recurrent Gaussian Splatting (RecGS), which takes advantage of today's photorealistic 3D reconstruction technology, 3DGS, to separate caustics from seafloor imagery. With a sequence of images taken by an underwater robot, we build 3DGS recurrently and decompose the caustic with low-pass filtering in each iteration. In the experiments, we analyze and compare with different methods, including joint optimization, 2D filtering, and deep learning approaches. The results show that our method can effectively separate the caustic from the seafloor, improving the visual appearance, and can be potentially applied on more problems with inconsistent illumination.

Gehan Zheng, Sanjay Seenivasan, Matthew Johnson-Roberson et al.

Imitation learning has enabled robots to acquire complex visuomotor manipulation skills from demonstrations, but deployment failures remain a major obstacle, especially for long-horizon action-chunked policies. Once execution drifts off the demonstration manifold, these policies often continue producing locally plausible actions without recovering from the failure. Existing runtime monitors either require failure data, over-trigger under benign feature drift, or stop at failure detection without providing a recovery mechanism. We present Rewind-IL, a training-free online safeguard framework for generative action-chunked imitation policies. Rewind-IL combines a zero-shot failure detector based on Temporal Inter-chunk Discrepancy Estimate (TIDE), calibrated with split conformal prediction, with a state-respawning mechanism that returns the robot to a semantically verified safe intermediate state. Offline, a vision-language model identifies recovery checkpoints in demonstrations, and the frozen policy encoder is used to construct a compact checkpoint feature database. Online, Rewind-IL monitors self-consistency in overlapping action chunks, tracks similarity to the checkpoint library, and, upon failure, rewinds execution to the latest verified safe state before restarting inference from a clean policy state. Experiments on real-world and simulated long-horizon manipulation tasks, including transfer to flow-matching action-chunked policies, demonstrate that policy-internal consistency coupled with semantically grounded respawning offers a practical route to improved reliability in imitation learning. Supplemental materials are available at https://sjay05.github.io/rewind-il

Weiming Zhi, Haozhan Tang, Tianyi Zhang et al.

Humans have the remarkable ability to use held objects as tools to interact with their environment. For this to occur, humans internally estimate how hand movements affect the object's movement. We wish to endow robots with this capability. We contribute methodology to jointly estimate the geometry and pose of objects grasped by a robot, from RGB images captured by an external camera. Notably, our method transforms the estimated geometry into the robot's coordinate frame, while not requiring the extrinsic parameters of the external camera to be calibrated. Our approach leverages 3D foundation models, large models pre-trained on huge datasets for 3D vision tasks, to produce initial estimates of the in-hand object. These initial estimations do not have physically correct scales and are in the camera's frame. Then, we formulate, and efficiently solve, a coordinate-alignment problem to recover accurate scales, along with a transformation of the objects to the coordinate frame of the robot. Forward kinematics mappings can subsequently be defined from the manipulator's joint angles to specified points on the object. These mappings enable the estimation of points on the held object at arbitrary configurations, enabling robot motion to be designed with respect to coordinates on the grasped objects. We empirically evaluate our approach on a robot manipulator holding a diverse set of real-world objects.

Chi Cuong Le, Weiming Zhi

Reliable robot perception requires not only predicting scene structure, but also identifying where predictions should be treated as unreliable due to sparse or missing observations. We present ContraMap, a contrastive continuous mapping method that augments kernel-based discriminative maps with an explicit uncertainty class trained using synthetic noise samples. This formulation treats unobserved regions as a contrastive class, enabling joint environment prediction and spatial uncertainty estimation in real time without Bayesian inference. Under a simple mixture-model view, we show that the probability assigned to the uncertainty class is a monotonic function of a distance-aware uncertainty surrogate. Experiments in 2D occupancy mapping, 3D semantic mapping, and tabletop scene reconstruction show that ContraMap preserves mapping quality, produces spatially coherent uncertainty estimates, and is substantially more efficient than Bayesian kernelmap baselines.

Ziyong Ma, Uksang Yoo, Jonathan Francis et al.

While soft robot manipulators offer compelling advantages over rigid counterparts, including inherent compliance, safe human-robot interaction, and the ability to conform to complex geometries, accurate forward modeling from low-dimensional actuation commands remains an open challenge due to nonlinear material phenomena such as hysteresis and manufacturing variability. We present SOFTMAP, a sim-to-real learning framework for real-time 3D forward modeling of tendon-actuated soft finger manipulators. SOFTMAP combines four components: (1) As-Rigid-As-Possible (ARAP)-based topological alignment that projects simulated and real point clouds into a shared, topologically consistent vertex space; (2) a lightweight MLP forward model pretrained on simulation data to map servo commands to full 3D finger geometry; (3) a residual correction network trained on a small set of real observations to predict per-vertex displacement fields that compensate for sim-to-real discrepancies; and (4) a closed-form linear actuation calibration layer enabling real-time inference at 30 FPS. We evaluate SOFTMAP on both simulated and physical hardware, achieving state-of-the-art shape prediction accuracy with a Chamfer distance of 0.389 mm in simulation and 3.786 mm on hardware, millimeter-level fingertip trajectory tracking across multiple target paths, and a 36.5% improvement in teleoperation task success over the baseline. Our results show that SOFTMAP provides a data-efficient approach for 3D forward modeling and control of soft manipulators.

Weiming Zhi

Contemporary robots have become exceptionally skilled at achieving specific tasks in structured environments. However, they often fail when faced with the limitless permutations of real-world unstructured environments. This motivates robotics methods which learn from experience, rather than follow a pre-defined set of rules. In this thesis, we present a range of learning-based methods aimed at enabling robots, operating in dynamic and unstructured environments, to better understand their surroundings, anticipate the actions of others, and take informed actions accordingly.

Haozhan Tang, Tianyi Zhang, Oliver Kroemer et al.

Robots operating in unstructured environments often require accurate and consistent object-level representations. This typically requires segmenting individual objects from the robot's surroundings. While recent large models such as Segment Anything (SAM) offer strong performance in 2D image segmentation. These advances do not translate directly to performance in the physical 3D world, where they often over-segment objects and fail to produce consistent mask correspondences across views. In this paper, we present GraphSeg, a framework for generating consistent 3D object segmentations from a sparse set of 2D images of the environment without any depth information. GraphSeg adds edges to graphs and constructs dual correspondence graphs: one from 2D pixel-level similarities and one from inferred 3D structure. We formulate segmentation as a problem of edge addition, then subsequent graph contraction, which merges multiple 2D masks into unified object-level segmentations. We can then leverage \emph{3D foundation models} to produce segmented 3D representations. GraphSeg achieves robust segmentation with significantly fewer images and greater accuracy than prior methods. We demonstrate state-of-the-art performance on tabletop scenes and show that GraphSeg enables improved performance on downstream robotic manipulation tasks. Code available at https://github.com/tomtang502/graphseg.git.

Tianyi Zhang, Kaining Huang, Weiming Zhi et al.

Humans have the remarkable ability to construct consistent mental models of an environment, even under limited or varying levels of illumination. We wish to endow robots with this same capability. In this paper, we tackle the challenge of constructing a photorealistic scene representation under poorly illuminated conditions and with a moving light source. We approach the task of modeling illumination as a learning problem, and utilize the developed illumination model to aid in scene reconstruction. We introduce an innovative framework that uses a data-driven approach, Neural Light Simulators (NeLiS), to model and calibrate the camera-light system. Furthermore, we present DarkGS, a method that applies NeLiS to create a relightable 3D Gaussian scene model capable of real-time, photorealistic rendering from novel viewpoints. We show the applicability and robustness of our proposed simulator and system in a variety of real-world environments.

Weiming Zhi, Haozhan Tang, Tianyi Zhang et al.

Representing the environment is a central challenge in robotics, and is essential for effective decision-making. Traditionally, before capturing images with a manipulator-mounted camera, users need to calibrate the camera using a specific external marker, such as a checkerboard or AprilTag. However, recent advances in computer vision have led to the development of \emph{3D foundation models}. These are large, pre-trained neural networks that can establish fast and accurate multi-view correspondences with very few images, even in the absence of rich visual features. This paper advocates for the integration of 3D foundation models into scene representation approaches for robotic systems equipped with manipulator-mounted RGB cameras. Specifically, we propose the Joint Calibration and Representation (JCR) method. JCR uses RGB images, captured by a manipulator-mounted camera, to simultaneously construct an environmental representation and calibrate the camera relative to the robot's end-effector, in the absence of specific calibration markers. The resulting 3D environment representation is aligned with the robot's coordinate frame and maintains physically accurate scales. We demonstrate that JCR can build effective scene representations using a low-cost RGB camera attached to a manipulator, without prior calibration.

Weiming Zhi, Ziyong Ma, Tianyi Zhang et al.

Autonomous robots typically need to construct representations of their surroundings and adapt their motions to the geometry of their environment. Here, we tackle the problem of constructing a policy model for collision-free motion generation, consistent with the environment, from a single input RGB image. Extracting 3D structures from a single image often involves monocular depth estimation. Developments in depth estimation have given rise to large pre-trained models such as DepthAnything. However, using outputs of these models for downstream motion generation is challenging due to frustum-shaped errors that arise. Instead, we propose a framework known as Video-Generation Environment Representation (VGER), which leverages the advances of large-scale video generation models to generate a moving camera video conditioned on the input image. Frames of this video, which form a multiview dataset, are then input into a pre-trained 3D foundation model to produce a dense point cloud. We then introduce a multi-scale noise approach to train an implicit representation of the environment structure and build a motion generation model that complies with the geometry of the representation. We extensively evaluate VGER over a diverse set of indoor and outdoor environments. We demonstrate its ability to produce smooth motions that account for the captured geometry of a scene, all from a single RGB input image.

Herbert Wright, Weiming Zhi, Martin Matak et al. · nvidia

Constructing 3D representations of object geometry is critical for many robotics tasks, particularly manipulation problems. These representations must be built from potentially noisy partial observations. In this work, we focus on the problem of reconstructing a multi-object scene from a single RGBD image using a fixed camera. Traditional scene representation methods generally cannot infer the geometry of unobserved regions of the objects in the image. Attempts have been made to leverage deep learning to train on a dataset of known objects and representations, and then generalize to new observations. However, this can be brittle to noisy real-world observations and objects not contained in the dataset, and do not provide well-calibrated reconstruction confidences. We propose BRRP, a reconstruction method that leverages preexisting mesh datasets to build an informative prior during robust probabilistic reconstruction. We introduce the concept of a retrieval-augmented prior, where we retrieve relevant components of our prior distribution from a database of objects during inference. The resulting prior enables estimation of the geometry of occluded portions of the in-scene objects. Our method produces a distribution over object shape that can be used for reconstruction and measuring uncertainty. We evaluate our method in both simulated scenes and in the real world. We demonstrate the robustness of our method against deep learning-only approaches while being more accurate than a method without an informative prior. Through real-world experiments, we particularly highlight the capability of BRRP to enable successful dexterous manipulation in clutter.

William Barron, Xiaoxiang Dong, Matthew Johnson-Roberson et al.

Teaching robots novel behaviors typically requires motion demonstrations via teleoperation or kinaesthetic teaching, that is, physically guiding the robot. While recent work has explored using human sketches to specify desired behaviors, data collection remains cumbersome, and demonstration datasets are difficult to scale. In this paper, we introduce an alternative paradigm, Learning from Cross-Modal Instructions, where robots are shaped by demonstrations in the form of rough annotations, which can contain free-form text labels, and are used in lieu of physical motion. We introduce the CrossInstruct framework, which integrates cross-modal instructions as examples into the context input to a foundational vision-language model (VLM). The VLM then iteratively queries a smaller, fine-tuned model, and synthesizes the desired motion over multiple 2D views. These are then subsequently fused into a coherent distribution over 3D motion trajectories in the robot's workspace. By incorporating the reasoning of the large VLM with a fine-grained pointing model, CrossInstruct produces executable robot behaviors that generalize beyond the environment of in the limited set of instruction examples. We then introduce a downstream reinforcement learning pipeline that leverages CrossInstruct outputs to efficiently learn policies to complete fine-grained tasks. We rigorously evaluate CrossInstruct on benchmark simulation tasks and real hardware, demonstrating effectiveness without additional fine-tuning and providing a strong initialization for policies subsequently refined via reinforcement learning.

Xiaoxiang Dong, Matthew Johnson-Roberson, Weiming Zhi

Learning from human video demonstrations offers a scalable alternative to teleoperation or kinesthetic teaching, but poses challenges for robot manipulators due to embodiment differences and joint feasibility constraints. We address this problem by proposing the Joint Flow Trajectory Optimization (JFTO) framework for grasp pose generation and object trajectory imitation under the video-based Learning-from-Demonstration (LfD) paradigm. Rather than directly imitating human hand motions, our method treats demonstrations as object-centric guides, balancing three objectives: (i) selecting a feasible grasp pose, (ii) generating object trajectories consistent with demonstrated motions, and (iii) ensuring collision-free execution within robot kinematics. To capture the multimodal nature of demonstrations, we extend flow matching to $\SE(3)$ for probabilistic modeling of object trajectories, enabling density-aware imitation that avoids mode collapse. The resulting optimization integrates grasp similarity, trajectory likelihood, and collision penalties into a unified differentiable objective. We validate our approach in both simulation and real-world experiments across diverse real-world manipulation tasks.

Wei-Teng Chu, Tianyi Zhang, Matthew Johnson-Roberson et al.

Implicit representations have been widely applied in robotics for obstacle avoidance and path planning. In this paper, we explore the problem of constructing an implicit distance representation from a single image. Past methods for implicit surface reconstruction, such as \emph{NeuS} and its variants generally require a large set of multi-view images as input, and require long training times. In this work, we propose Fast Image-to-Neural Surface (FINS), a lightweight framework that can reconstruct high-fidelity surfaces and SDF fields based on a single or a small set of images. FINS integrates a multi-resolution hash grid encoder with lightweight geometry and color heads, making the training via an approximate second-order optimizer highly efficient and capable of converging within a few seconds. Additionally, we achieve the construction of a neural surface requiring only a single RGB image, by leveraging pre-trained foundation models to estimate the geometry inherent in the image. Our experiments demonstrate that under the same conditions, our method outperforms state-of-the-art baselines in both convergence speed and accuracy on surface reconstruction and SDF field estimation. Moreover, we demonstrate the applicability of FINS for robot surface following tasks and show its scalability to a variety of benchmark datasets.

Haisheng Wang, Weiming Zhi

This paper formulates the Embodied Questions Answering (EQsA) problem, introduces a corresponding benchmark, and proposes a system to tackle the problem. Classical Embodied Question Answering (EQA) is typically formulated as answering one single question by actively exploring a 3D environment. Real deployments, however, often demand handling multiple questions that may arrive asynchronously and carry different urgencies. We formalize this setting as Embodied Questions Answering (EQsA) and present ParaEQsA, a framework for parallel, urgency-aware scheduling and answering. ParaEQsA leverages a group memory module shared among questions to reduce redundant exploration, and a priority-planning module to dynamically schedule questions. To evaluate this setting, we contribute the Parallel Asynchronous Embodied Questions (PAEQs) benchmark containing 40 indoor scenes and five questions per scene (200 in total), featuring asynchronous follow-up questions and urgency labels. We further propose metrics for EQsA performance: Direct Answer Rate (DAR), and Normalized Urgency-Weighted Latency (NUWL), which jointly measure efficiency and responsiveness of this system. ParaEQsA consistently outperforms strong sequential baselines adapted from recent EQA systems, while reducing exploration and delay. Empirical evaluations investigate the relative contributions of priority, urgency modeling, spatial scope, reward estimation, and dependency reasoning within our framework. Together, these results demonstrate that urgency-aware, parallel scheduling is key to making embodied agents responsive and efficient under realistic, multi-question workloads.

Haozhan Tang, Tianyi Zhang, Matthew Johnson-Roberson et al.

Robot manipulation, especially bimanual manipulation, often requires setting up multiple cameras on multiple robot manipulators. Before robot manipulators can generate motion or even build representations of their environments, the cameras rigidly mounted to the robot need to be calibrated. Camera calibration is a cumbersome process involving collecting a set of images, with each capturing a pre-determined marker. In this work, we introduce the Bi-Manual Joint Calibration and Representation Framework (Bi-JCR). Bi-JCR enables multiple robot manipulators, each with cameras mounted, to circumvent taking images of calibration markers. By leveraging 3D foundation models for dense, marker-free multi-view correspondence, Bi-JCR jointly estimates: (i) the extrinsic transformation from each camera to its end-effector, (ii) the inter-arm relative poses between manipulators, and (iii) a unified, scale-consistent 3D representation of the shared workspace, all from the same captured RGB image sets. The representation, jointly constructed from images captured by cameras on both manipulators, lives in a common coordinate frame and supports collision checking and semantic segmentation to facilitate downstream bimanual coordination tasks. We empirically evaluate the robustness of Bi-JCR on a variety of tabletop environments, and demonstrate its applicability on a variety of downstream tasks.

Tianyi Zhang, Weiming Zhi, Joshua Mangelson et al.

This paper tackles the problem of generating representations of underwater 3D terrain. Off-the-shelf generative models, trained on Internet-scale data but not on specialized underwater images, exhibit downgraded realism, as images of the seafloor are relatively uncommon. To this end, we introduce DreamSea, a generative model to generate hyper-realistic underwater scenes. DreamSea is trained on real-world image databases collected from underwater robot surveys. Images from these surveys contain massive real seafloor observations and covering large areas, but are prone to noise and artifacts from the real world. We extract 3D geometry and semantics from the data with visual foundation models, and train a diffusion model that generates realistic seafloor images in RGBD channels, conditioned on novel fractal distribution-based latent embeddings. We then fuse the generated images into a 3D map, building a 3DGS model supervised by 2D diffusion priors which allows photorealistic novel view rendering. DreamSea is rigorously evaluated, demonstrating the ability to robustly generate large-scale underwater scenes that are consistent, diverse, and photorealistic. Our work drives impact in multiple domains, spanning filming, gaming, and robot simulation.

Hongzhe Cheng, Tianyou Zheng, Tianyi Zhang et al.

Out-of-Distribution(OOD) detection, a fundamental machine learning task aimed at identifying abnormal samples, traditionally requires model retraining for different inlier distributions. While recent research demonstrates the applicability of diffusion models to OOD detection, existing approaches are limited to Euclidean or latent image spaces. Our work extends OOD detection to trajectories in the Special Euclidean Group in 3D ($\mathbb{SE}(3)$), addressing a critical need in computer vision, robotics, and engineering applications that process object pose sequences in $\mathbb{SE}(3)$. We present $\textbf{D}$iffusion-based $\textbf{O}$ut-of-distribution detection on $\mathbb{SE}(3)$ ($\mathbf{DOSE3}$), a novel OOD framework that extends diffusion to a unified sample space of $\mathbb{SE}(3)$ pose sequences. Through extensive validation on multiple benchmark datasets, we demonstrate $\mathbf{DOSE3}$'s superior performance compared to state-of-the-art OOD detection frameworks.

Tin Lai, Weiming Zhi, Tucker Hermans et al.

We propose Parallelised Diffeomorphic Sampling-based Motion Planning (PDMP). PDMP is a novel parallelised framework that uses bijective and differentiable mappings, or diffeomorphisms, to transform sampling distributions of sampling-based motion planners, in a manner akin to normalising flows. Unlike normalising flow models which use invertible neural network structures to represent these diffeomorphisms, we develop them from gradient information of desired costs, and encode desirable behaviour, such as obstacle avoidance. These transformed sampling distributions can then be used for sampling-based motion planning. A particular example is when we wish to imbue the sampling distribution with knowledge of the environment geometry, such that drawn samples are less prone to be in collisions. To this end, we propose to learn a continuous occupancy representation from environment occupancy data, such that gradients of the representation defines a valid diffeomorphism and is amenable to fast parallel evaluation. We use this to "morph" the sampling distribution to draw far fewer collision-prone samples. PDMP is able to leverage gradient information of costs, to inject specifications, in a manner similar to optimisation-based motion planning methods, but relies on drawing from a sampling distribution, retaining the tendency to find more global solutions, thereby bridging the gap between trajectory optimisation and sampling-based planning methods.

Weiming Zhi, Lionel Ott, Fabio Ramos

This work addresses the problem of predicting the motion trajectories of dynamic objects in the environment. Recent advances in predicting motion patterns often rely on machine learning techniques to extrapolate motion patterns from observed trajectories, with no mechanism to directly incorporate known rules. We propose a novel framework, which combines probabilistic learning and constrained trajectory optimisation. The learning component of our framework provides a distribution over future motion trajectories conditioned on observed past coordinates. This distribution is then used as a prior to a constrained optimisation problem which enforces chance constraints on the trajectory distribution. This results in constraint-compliant trajectory distributions which closely resemble the prior. In particular, we focus our investigation on collision constraints, such that extrapolated future trajectory distributions conform to the environment structure. We empirically demonstrate on real-world and simulated datasets the ability of our framework to learn complex probabilistic motion trajectories for motion data, while directly enforcing constraints to improve generalisability, producing more robust and higher quality trajectory distributions.

Weiming Zhi, Tin Lai, Lionel Ott et al.

Advances in differentiable numerical integrators have enabled the use of gradient descent techniques to learn ordinary differential equations (ODEs). In the context of machine learning, differentiable solvers are central for Neural ODEs (NODEs), a class of deep learning models with continuous depth, rather than discrete layers. However, these integrators can be unsatisfactorily slow and inaccurate when learning systems of ODEs from long sequences, or when solutions of the system vary at widely different timescales in each dimension. In this paper we propose an alternative approach to learning ODEs from data: we represent the underlying ODE as a vector field that is related to another base vector field by a differentiable bijection, modelled by an invertible neural network. By restricting the base ODE to be amenable to integration, we can drastically speed up and improve the robustness of integration. We demonstrate the efficacy of our method in training and evaluating continuous neural networks models, as well as in learning benchmark ODE systems. We observe improvements of up to two orders of magnitude when integrating learned ODEs with GPUs computation.

Weiming Zhi, Tin Lai, Lionel Ott et al.

Critical for the coexistence of humans and robots in dynamic environments is the capability for agents to understand each other's actions, and anticipate their movements. This paper presents Stochastic Process Anticipatory Navigation (SPAN), a framework that enables nonholonomic robots to navigate in environments with crowds, while anticipating and accounting for the motion patterns of pedestrians. To this end, we learn a predictive model to predict continuous-time stochastic processes to model future movement of pedestrians. Anticipated pedestrian positions are used to conduct chance constrained collision-checking, and are incorporated into a time-to-collision control problem. An occupancy map is also integrated to allow for probabilistic collision-checking with static obstacles. We demonstrate the capability of SPAN in crowded simulation environments, as well as with a real-world pedestrian dataset.

Weiming Zhi, Tin Lai, Lionel Ott et al.

Being able to safely operate for extended periods of time in dynamic environments is a critical capability for autonomous systems. This generally involves the prediction and understanding of motion patterns of dynamic entities, such as vehicles and people, in the surroundings. Many motion prediction methods in the literature can learn a function, mapping position and time to potential trajectories taken by people or other dynamic entities. However, these predictions depend only on previously observed trajectories, and do not explicitly take into consideration the environment. Trends of motion obtained in one environment are typically specific to that environment, and are not used to better predict motion in other environments. In this paper, we address the problem of generating likely motion dynamics conditioned on the environment, represented as an occupancy map. We introduce the Occupancy Conditional Trajectory Network (OCTNet) framework, capable of generalising the previously observed motion in known environments, to generate trajectories in new environments where no observations of motion has not been observed. OCTNet encodes trajectories as a fixed-sized vector of parameters and utilises neural networks to learn conditional distributions over parameters. We empirically demonstrate our method's ability to generate complex multi-modal trajectory patterns in different environments.

Tin Lai, Weiming Zhi, Fabio Ramos

Trajectory modelling had been the principal research area for understanding and anticipating human behaviour. Predicting the dynamic path by observing the agent and its surrounding environment are essential for applications such as autonomous driving and indoor navigation suggestions. However, despite the numerous researches that had been presented, most available dataset does not contains any information on environmental factors---such as the occupancy representation of the map---which arguably plays a significant role on how an agent chooses its trajectory. We present a trajectory dataset with the corresponding occupancy representations of different local-maps. The dataset contains more than 120 locally-structured maps with occupancy representation and more than 110K trajectories in total. Each map has few hundred corresponding simulated trajectories that navigate from a spatial location of a room to another point. The dataset is freely available online.

Weiming Zhi, Lionel Ott, Fabio Ramos

Understanding the dynamics of an environment, such as the movement of humans and vehicles, is crucial for agents to achieve long-term autonomy in urban environments. This requires the development of methods to capture the multi-modal and probabilistic nature of motion patterns. We present Kernel Trajectory Maps (KTM) to capture the trajectories of movement in an environment. KTMs leverage the expressiveness of kernels from non-parametric modelling by projecting input trajectories onto a set of representative trajectories, to condition on a sequence of observed waypoint coordinates, and predict a multi-modal distribution over possible future trajectories. The output is a mixture of continuous stochastic processes, where each realisation is a continuous functional trajectory, which can be queried at arbitrarily fine time steps.