Publications

ContactGuard predicts future visual latents from a policy's planned actions to abort likely contact failures before they happen, without modifying the policy.

Depth foundation models pick different layers in transparent scenes, and a training-free Laplacian input transform can shift which layer some models report.

Rewind-IL detects when an imitation-learning policy is failing using its own inconsistency, then rewinds the robot to a verified safe state and restarts.

Robots can learn motions from rough image annotations and text instead of physical demonstrations, using a vision-language model whose 2D outputs are fused into 3D trajectories.

FINS builds a neural signed distance field from a single image in about ten seconds, enabling robot surface-following motion.

A framework turns human videos into feasible robot grasps and object trajectories by jointly optimizing grasp, motion likelihood, and collision avoidance.

NS-MAE is a single self-supervised pretraining method for LiDAR and camera data that uses NeRF-style rendering to improve driving perception with few labels.

Bi-JCR calibrates cameras on two robot arms and builds a shared 3D scene model from ordinary images, with no calibration markers.

VGER turns one RGB image into a generated camera-motion video, then a 3D point cloud, to produce smooth collision-free robot motions.

GraphSeg merges over-segmented 2D masks from a few RGB images into consistent 3D object segments using graph edge addition and contraction.

DreamSea trains a diffusion model on robot-survey images to generate large, consistent, photorealistic 3D underwater terrain rendered with Gaussian splatting.

A training-free Laplacian image prompt lets pretrained depth models reveal a second depth layer behind glass, evaluated on the new MD-3k benchmark.

DOSE3 detects out-of-distribution robot and vehicle pose trajectories using a diffusion model over SE(3), without retraining when the inlier set changes.

A noise-synthesis pipeline, the Robust-Ego3D benchmark, and CorrGS test-time adaptation make ego-motion and 3D reconstruction hold up on noisy video.

A single user sketch over a camera image teaches a robot to approach a surface and follow a stable, repeating motion.

A robot recovers the shape, scale, and pose of a grasped object from uncalibrated RGB images by aligning 3D foundation model estimates to its own frame.

JCR uses 3D foundation models to build a scene representation and calibrate a wrist-mounted RGB camera at once, without calibration markers.

RecGS repeatedly fits a 3D Gaussian Splatting model and filters the residual to strip water caustics from underwater robot imagery of the seafloor.

A learned model of a robot's moving onboard light lets 3D Gaussian scenes be reconstructed, relit and rendered in real time in the dark.

V-PRISM builds probabilistic 3D segmentation maps of tabletop scenes from one segmented RGB-D image, giving object shapes and uncertainty.

Users teach robots new skills by sketching trajectories on 2D camera images, which are converted into a probabilistic model of 3D motion.

Projecting encoder embeddings onto a unit hypersphere makes point cloud completion training more stable and improves results in single-task and multi-task learning.

A neural reflectance field combined with a physical water model restores true color in underwater images, with better color consistency than existing methods.

An unsupervised convolutional network removes snow from LiDAR point clouds 52x faster than prior methods, with better de-noising and usable mapping.

CLONeR splits NeRF into LiDAR-trained occupancy and camera-trained color, and adds an occupancy grid, to model sparse outdoor driving scenes.

A two-stage controller for underwater vehicles picks energy-minimizing speeds and pitch under ocean currents, saving over 13% energy in simulated survey and inspection missions.

Overhead cameras, CNN detectors and Kalman filtering track zoo dolphins over about 100 hours, revealing habitat use and swimming patterns that manual annotation cannot capture.

An economic model predictive controller cuts underwater vehicle energy use during waypoint tracking, nearly matching offline optimal trajectories and beating a line-of-sight baseline in simulation.

An underwater vehicle with only an IMU and depth sensor moves deliberately to estimate water flow, which in turn improves its own navigation.

A vision-based system lets underwater ROV manipulators automatically collect and return biological samples, demonstrated in field trials including inside the active Kolumbo volcano.

A visual SLAM system combines a vehicle stereo camera with a wrist-mounted fisheye camera to map natural seafloors for underwater manipulators.

A self-supervised method locates camera images on small-scale multi-modal maps in real time, correcting drift for robots in GPS-denied environments.

An underwater vehicle plans paths that both save energy and gather flow measurements to reduce current-prediction uncertainty, showing energy savings in simulation.

A multi-task network links pedestrian intent and actions, using predicted future actions to improve crossing intent detection and action prediction on PIE and JAAD.

A kinematic model of car-following and lane changes predicts highway trajectories in general scenarios, staying interpretable while outperforming state-of-the-art methods on highway datasets.

A point cloud network that learns features in local reference frames stays robust to rotation without needing rotation-based data augmentation.

BiTraP predicts pedestrian paths by estimating a goal and decoding bi-directionally, reducing long-horizon error by about 10\u201350% over prior methods.

A probabilistic car-following model predicts nearby vehicles' trajectories from just 400 ms of observation, matching state-of-the-art accuracy on dense NGSIM highway traffic.

Local point sets each vote on a complete shape's latent feature, making classification, segmentation and completion more robust to partial point clouds.

By descending in a spiral, an underwater vehicle with only inertial and depth sensors can estimate ocean currents and localize better than dead-reckoning in simulation.

An underwater vehicle controller picks forward speeds that cut energy use on a planned path while staying robust to errors in ocean current estimates.

Thermal camera blur comes from pixel thermal inertia, so deblurring each pixel over time with a LASSO formulation outperforms spatial deblurring methods.

A real-time, interpretable model learns when pedestrians yield to vehicles and how that changes their motion, predicting trajectories in shared spaces.

Calibrates an underwater camera inside a spherical dome by using the point spread function and measured distortion to recover its pose, even with blur.

An open-source tool lets engineers explore underwater camera system designs using physical light models, replacing trial-and-error selection of cameras, lenses, housings, and settings.

Neural networks classify 3D coral reef reconstructions by combining multiple image views per point, reaching about 96% overall accuracy.

A predictive coding network forecasts pedestrian poses from 2D camera video without labels, tested on two real-world autonomous vehicle data sets.

ARMTD plans robot arm trajectories in real time using precomputed reachable sets and a braking fail-safe, avoiding collisions in simulation and hardware tests.

Dynamic Watermarking is extended to time-varying vehicle dynamics, detecting replay attacks in a CarSim car model and a 1/10 scale rover.

SilhoNet is adapted to fisheye images by projecting each region of interest onto a tangent plane, and an underwater fisheye pose dataset, UWHandles, is released.

A low-cost pipeline combining a 2D planar LiDAR and one camera detects pedestrians and their orientation about as well as 3D-LiDAR methods.

RTD-Interval computes reachable sets over time intervals, removing obstacle buffering so vehicles plan faster and less conservatively while avoiding at-fault collisions.

Reachability analysis on a simple template model, with error bounds to the full model, lets online model predictive control choose bipedal gaits that stay safe.

Dynamic watermarking is extended from time-invariant to time-varying systems, detecting generalized replay attacks on a simulated vehicle model.

Combining forward and backward reachability lets a self-driving car assess occlusion risk efficiently, driving faster and cutting median collisions 7.5× at occluded intersections.

An economic model predictive controller for underwater vehicles minimizes energy use by pairing a short-horizon cost with an energy-to-go estimate for reaching the goal.

An unsupervised network turns raw underwater stereo images into dense depth maps and color-corrected images, using image formation physics and stereo geometry instead of ground truth.

LiStereo fuses sparse LIDAR with stereo images to produce dense depth maps, trained without ground-truth labels and robust to low-resolution LIDAR.

A particle filter fuses on-body motion tags with overhead cameras to localize and track free-swimming dolphins, even when camera views are sparse.

A simulator seeded with a little real data generates synthetic pedestrian trajectories that train Social GAN to outperform training on real annotated data.

Reachability-based Trajectory Design is applied to a full-dynamics passenger car in simulation, planning safe trajectories in real time where NMPC and RRT comparisons fall short.

A micro-turbine speed sensor on animal tags was calibrated in flumes and towing tanks, and matched video-based speed and distance for dolphins.

A method sets Dynamic Watermarking attack detectors to a chosen false alarm rate and measures attacker capability, tested against classic detectors in simulation and on a Segway.

A computationally efficient controller-planner for redundant arms keeps joints within limits, and was tested on a real 7-DOF manipulator.

Real underwater images draped over randomized rendered terrain yield stereo training pairs with known depth, improving underwater disparity estimation over classical and terrestrial-trained methods.

An open source AUV under 1 m long and 20 kg, rated to 6000 m, is designed for deep-sea optical mapping at under $30,000 in materials.

Reachability-based Trajectory Design precomputes a robot's tracking-error-bounded reachable set so real-time receding-horizon planning stays safe and always has a next trajectory ready.

SilhoNet predicts object silhouettes from a single RGB image and regresses 6D pose from them, outperforming two monocular baselines on YCB-Video.

Learning camera-specific effects like blur, noise, and color temperature from real data and applying them to synthetic images reduces the sim-to-real gap in driving detection.

A single stereo network refines disparity using features learned for semantic segmentation, training without ground-truth depth and outputting both depth and labels.

A two-stage model predictive controller with switching logic gets near-optimal energy use for point-to-point AUV travel in simulation, at much lower computational cost.

Using road maps to estimate risk from unseen vehicles lets autonomous cars plan safer, smoother paths, cutting simulated collisions 4.8 times versus a baseline.

A recurrent network trained with gait periodicity, body symmetry, and ground-force cues predicts multi-pedestrian 3D body pose and location at urban intersections.

PedX is a stereo and LiDAR dataset of pedestrians at urban intersections, with metric 3D body models produced by a multimodal fitting method.

A physically based pipeline that simulates camera effects in training images makes vehicle detectors more robust and better at generalizing across datasets.

A real-time sonar-based SLAM system uses learned loop closure proposals and constraint-based data association to localize underwater robots during ship hull inspection.

A Bayesian, decentralized deployment method lets robot teams keep mapping an environment accurately even when some robots' sensors fail.

Dolphins are localized by fusing automatically detected video positions with wearable inertial sensor data through Kalman-style filters and iSAM2 pose-graph smoothing.

A quantitative benchmark of how well in-air 3D reconstruction methods work underwater, including a comparison of low-cost and high-cost cameras.

Intrinsic image decomposition for RGB-D video is cast as recursive Bayesian filtering, giving more temporally consistent results than frame-by-frame optimization.

A light field camera lets underwater images be dehazed by combining per-view 2D dehazing with guided filtering, tested on a public water-tank dataset.

Inconsistencies in a detector's output across consecutive or stereo images can flag missed detections automatically, without human labels.

Pedestrian motion is forecast probabilistically using weighted sums of learned differential equations, running in real time and beating prior methods over long horizons.

A low-cost one-way travel-time acoustic ranging system using surface buoys could localize deep-sea micro AUVs, cutting simulated error by a factor of three.

Attenuation coefficients for underwater color correction can be estimated automatically inside bundle adjustment, at the same time as 3D structure is recovered.

Planning with a simple model plus a precomputed reachable set of the true vehicle's motion lets runtime obstacle checks guarantee collision-free trajectories.

WaterGAN synthesizes realistic underwater images from in-air RGB-D data to train a network that color-corrects single underwater images and estimates coarse depth.

A vehicle detector trained only on automatically labeled video-game images outperformed the same network trained on human-annotated real data when tested on KITTI.

A plenoptic camera and GPU pipeline produce accurate, real-time dense 3D reconstructions of objects underwater, with correction for light attenuation.

Deep features are compressed into a space linear in the state, with perturbation-based uncertainty, enabling real-time Kalman filtering for sonar-based underwater localization.

An AUV and a diver-held stereo rig produced a 26,600 m2, 2 mm/pixel 3D map of the submerged city of Pavlopetri.

Using 3D structure from images and an underwater image formation model, this method recovers true seafloor color from AUV imagery without reflectance panels.

An autonomous surface vessel builds textured 3D seafloor maps of shallow areas by projecting color-corrected camera images onto a sonar-derived mesh.

An automated pipeline finds foundation rocks and walls in stereo seafloor maps of a submerged city and outputs geo-referenced contours for archaeologists.

Sensor fusion is framed as a prediction market where algorithms trade, and the price equilibrium matches hand-designed fusion on KITTI pedestrian detection.

Matching salient image patches instead of points lets underwater robots recognize places across years despite lighting changes and biofouling.

Logged touch interactions on a mobile app for underwater 3D maps can be mined to find salient regions, validated against eye-tracking data.

An AUV planner combines sonar for coarse mapping and stereo cameras for detail, estimating seafloor rugosity faster than other survey planning methods in simulation.

Touchscreen interaction logs from people exploring underwater 3D maps can be mined to find salient points, validated against eye-tracking data.

Three automatic methods align seafloor survey maps taken up to two years apart, enabling long-term monitoring of reefs and kelp forests.

A cheap kite-mounted camera and structure-from-motion pipeline produce sub-centimeter 3D multi-spectral maps of a 200 m intertidal rocky shore.

Treating seafloor elevation maps as signals, this method separates ground from objects by frequency cut-off, producing segments useful for navigation across varied terrains.

An out-of-core texture blending method turns underwater stereo images into high-resolution, georeferenced 3D seafloor maps without exceeding memory limits.

A diver-propelled stereo camera system, using robotic mapping methods, produced 2D mosaics and 3D textured models of the submerged Bronze Age town of Pavlopetri.

Seafloor rugosity, slope and aspect can be computed at many scales from stereo image 3D reconstructions, replacing slower, more invasive in-situ measurements.

A kite-borne consumer camera and automated photogrammetry produce sub-centimeter 3D, colour maps of intertidal rock platforms cheaply, demonstrated at Cape Banks, Sydney.

Using 3D structure from structure-from-motion, the method corrects underwater attenuation and lighting effects to produce colour-consistent photo-textured models from AUV imagery.

Segmenting seafloor scans by separating low-frequency ground from objects in the Fourier domain yields object features that support navigation and scan alignment.

An Australia-wide program uses autonomous underwater vehicles to repeatedly image reference sites on the continental shelf, building baselines for tracking long-term seafloor change.

A robot completes the hidden side of unknown tabletop objects by fitting symmetries, then plans stable grasps on the completed shape.

A color-and-texture saliency method ranks AUV seafloor images so reviewers can find unusual ones, and it agrees well with human labels.

A min-cut method labels multiple objects in a 3D point cloud at once, with several seeding strategies compared on real robot-head data.

An automated pipeline turns underwater robot stereo images and SLAM poses into textured 3D seafloor models that scientists can explore during research cruises.

A drone fuses IMU, GPS, and camera data with least-squares smoothing to estimate its path and build textured 3D terrain maps.

A method merges ship-based sonar with AUV sonar and stereo images into a single seafloor model spanning kilometers and fine detail.

Stereo SLAM and dense reconstruction tools let the AUV Sirius build detailed seafloor models, tested on 2007 reef surveys at Ningaloo and the Great Barrier Reef.

An efficient visual SLAM method uses stereo loop closures and Cholesky updates to correct underwater vehicle trajectories during large sponge bed surveys.

A pipeline using spin image feature matching registers range data from two AUV views into visually consistent 3D seafloor meshes.

Adding stereo-derived surface shape and roughness to a 2D pipeline improves coral classification and moves toward real-time underwater survey processing.

A two-step method combines camera and acoustic data from underwater vehicles to segment and classify coral, producing coral density statistics during surveys.