Perception under incomplete observations
Shape completion, uncertainty and persistent mapping.
Research
My work studies how priors and partial observations can guide robotic perception, search and manipulation.
Research themes
Structured reasoning and active perception
Scene graphs, open-vocabulary object search and view planning.
Perception–action integration
Mobile manipulation, multi-arm systems and shared autonomy.
Featured projects
SG-AMP: Scene-Graph-Guided Active Perception and Semantics-Aware Motion Planning for Pepper Plants
IROS 2026 Workshop · Pittsburgh, USA
SG-AMP connects uncertainty-aware depth completion, persistent panoptic mapping, plant scene graphs and view-motion planning for inspecting occluded crops. It achieved 55.27% semantic mIoU, 38.67% PQ and 40.62 mm depth RMSE on pepper data.
SG-DOR: Learning Scene Graphs with Direction-Conditioned Occlusion Reasoning for Pepper Plants
IROS 2026 · Pittsburgh, USA
SG-DOR uses direction-conditioned scene graphs to infer plant-part attachments and rank leaves that obstruct a pepper from a chosen approach direction. It achieved 0.73 occlusion F1 and 0.85 NDCG@3.
O-STaR: Open-Vocabulary Object Search through Spatio-Temporal Reasoning on Dynamic Scene Graphs
GRC 2026 · Cologne, Germany
O-STaR searches for relocated and concealed objects by grounding language-model priors in scene geometry and updating beliefs on a dynamic scene graph. It was validated on Stretch and in 60-day simulated scenarios; an extended version is under review at ICRA 2027.
EvidMTL: Evidential Multi-Task Learning for Uncertainty-Aware Semantic Surface Mapping from Monocular RGB Images
IROS 2025 · Hangzhou, China
EvidMTL introduces evidential depth and semantic heads with a new evidential depth loss that calibrates predictive belief strength from monocular RGB. EvidKimera uses these uncertainties for consistent 3D mapping and outperformed Kimera by 30% in zero-shot ScanNetV2 semantic surface mapping accuracy and consistency.
Open-Vocabulary and Semantic-Aware Reasoning for Search and Retrieval of Objects in Dynamic and Concealed Spaces
IROS 2025 Workshop · Hangzhou, China · Best Poster
Stretch-Compose combines spatial, semantic and geometric reasoning to retrieve relocated or concealed objects from an outdated map. Real-robot trials achieved 100% navigation and open-space detection while reducing concealed-space search time by 68%.
HortiBot: An Adaptive Multi-Arm System for Robotic Horticulture of Sweet Peppers
IROS 2024 · Abu Dhabi, UAE
HortiBot’s perception stack performs real-time pepper, peduncle and stem detection, online data association, and scene-graph world modelling of pepper plants. The integrated three-arm system uses active perception and dual-arm manipulation for selective harvesting of real peppers.
DawnIK: Decentralized Collision-Aware Inverse Kinematics Solver for Heterogeneous Multi-Arm Systems
Humanoids 2023 · Austin, USA
DawnIK expresses inverse-kinematics and collision constraints as weighted costs in a nonlinear optimizer, without requiring inverse-Jacobian computation or redundant degrees of freedom. Against CollisionIK, it produced no collisions in any evaluated multi-arm scenario and improved trajectory tracking when collision risk was low.
NBV-SC: Next Best View Planning Based on Shape Completion for Fruit Mapping and Reconstruction
IROS 2023 · Detroit, USA
NBV-SC synthesizes targeted views of predicted missing fruit surfaces and rejects redundant samples using viewpoint dissimilarity. It improved fruit-size estimation and reconstruction, reduced viewpoint-sampling time by 80–90% in the harder simulated scenarios, and was demonstrated with a UR5e in a commercial glasshouse.
Fruit Mapping with Shape Completion for Autonomous Crop Monitoring
CASE 2022 · Mexico City, Mexico
Fruit Mapping fuses segmented RGB-D observations into a real-time TSDF map, clusters the reconstructed fruit surfaces, and fits optimized superellipsoids to estimate complete shapes and volumes despite occlusions. It accurately estimated fruit volumes in simulated multi-plant scenarios and recovered pepper shapes from data recorded in a commercial glasshouse.
Flexible, Semi-Autonomous Grasping for Assistive Robotics
ICRA 2016 · Stockholm, Sweden
The system combines a dense graspability map with dynamic virtual fixtures that adapt online to the operator’s commanded motion: the user controls translation while the system orients the hand and automates grasp closure. In a SpaceMouse user study, the semi-autonomous mode reduced task-completion time and subjective workload compared with pure teleoperation.