Projects | MoniJesu James, Robotics Researcher & Engineer

Research & Engineering

Featured Projects

Research in navigation, semantic mapping and robot coordination, alongside practical robotics systems.

AnyGoal

Training-free lifelong navigation with language, image and object goals.

Problem: Robot teams must search unfamiliar scenes and remember useful evidence across changing goals without task-specific retraining.
My contribution: I formulated the approach, implemented the system, designed and ran the experiments, and wrote the manuscript. A persistent Bayesian Value Map combines uncertain vision-language evidence across views and subtasks to guide exploration and agent allocation.
Result: On GOAT-Bench val_unseen (360 episodes, 2,669 subtasks), two agents achieved 52.4% subtask success versus 41.9% for one agent. First-author arXiv preprint.

Vision-language navigationBayesian mappingMulti-agent explorationGOAT-Bench

Read paper ↗

GoalVLM

Cooperative open-vocabulary object navigation using vision-language reasoning.

Problem: Agents need to interpret new object goals and choose where to explore without a fixed target vocabulary.
My contribution: I led the research, implementation and evaluation: VLM-guided frontier reasoning, text-prompted perception, depth-projected BEV semantic mapping and calibrated goal localization.
Result: Two agents reached 55.8% subtask success and 18.3% SPL over 1,032 sequential object-goal subtasks on GOAT-Bench val_unseen, with no task-specific training. First-author arXiv preprint.

VLM reasoningSAM3BEV mappingHabitat

Read paper ↗

GoalSwarm

Decentralized semantic coordination for aerial robot teams.

Problem: UAV teams searching for open-vocabulary targets need to share semantic evidence and coordinate exploration without duplicating coverage.
My contribution: I developed decentralized multi-UAV coordination using semantic frontiers, geodesic cost-utility bidding and spatial separation penalties, supported by shared lightweight top-down semantic maps.
Result: A decentralized framework combining shared semantic maps, Bayesian goal evidence and coordinated frontier allocation for aerial object navigation. First-author arXiv preprint.

Multi-UAV systemsSemantic mappingFrontier allocationOpen-vocabulary perception

Read paper ↗

Skyliner

Current research on reliable vision-language navigation for GPS-denied UAVs.

Problem: Visual guidance can be useful but unreliable; navigation must retain a dependable route-control fallback when landmarks cannot be grounded.
My contribution: As a Research Technician at Skoltech, I co-develop arbitration between a structured route prior and visual guidance, with landmark verification and persistent evidence before arrival is declared.
Result: Demonstrated closed-loop route-to-vision-to-route transitions. Ongoing research; no public paper or code release is linked.

UAV navigationVisual groundingControl arbitrationEvidence-gated stopping

Drone Human Object Detection

ROS2 package for real-time detection and 3D localization of humans and objects using drone-mounted RealSense camera with GroundingDINO zero-shot detection.

ROS2GroundingDINORealSensePythonDeep Learning

View on GitHub

Research & Engineering

Featured Projects

Research in navigation, semantic mapping and robot coordination, alongside practical robotics systems.

AnyGoal

Training-free lifelong navigation with language, image and object goals.

Problem: Robot teams must search unfamiliar scenes and remember useful evidence across changing goals without task-specific retraining.
My contribution: I formulated the approach, implemented the system, designed and ran the experiments, and wrote the manuscript. A persistent Bayesian Value Map combines uncertain vision-language evidence across views and subtasks to guide exploration and agent allocation.
Result: On GOAT-Bench val_unseen (360 episodes, 2,669 subtasks), two agents achieved 52.4% subtask success versus 41.9% for one agent. First-author arXiv preprint.

Vision-language navigationBayesian mappingMulti-agent explorationGOAT-Bench

Read paper ↗

GoalVLM

Cooperative open-vocabulary object navigation using vision-language reasoning.

Problem: Agents need to interpret new object goals and choose where to explore without a fixed target vocabulary.
My contribution: I led the research, implementation and evaluation: VLM-guided frontier reasoning, text-prompted perception, depth-projected BEV semantic mapping and calibrated goal localization.
Result: Two agents reached 55.8% subtask success and 18.3% SPL over 1,032 sequential object-goal subtasks on GOAT-Bench val_unseen, with no task-specific training. First-author arXiv preprint.

VLM reasoningSAM3BEV mappingHabitat

Read paper ↗

GoalSwarm

Decentralized semantic coordination for aerial robot teams.

Problem: UAV teams searching for open-vocabulary targets need to share semantic evidence and coordinate exploration without duplicating coverage.
My contribution: I developed decentralized multi-UAV coordination using semantic frontiers, geodesic cost-utility bidding and spatial separation penalties, supported by shared lightweight top-down semantic maps.
Result: A decentralized framework combining shared semantic maps, Bayesian goal evidence and coordinated frontier allocation for aerial object navigation. First-author arXiv preprint.

Multi-UAV systemsSemantic mappingFrontier allocationOpen-vocabulary perception

Read paper ↗

Skyliner

Current research on reliable vision-language navigation for GPS-denied UAVs.

Problem: Visual guidance can be useful but unreliable; navigation must retain a dependable route-control fallback when landmarks cannot be grounded.
My contribution: As a Research Technician at Skoltech, I co-develop arbitration between a structured route prior and visual guidance, with landmark verification and persistent evidence before arrival is declared.
Result: Demonstrated closed-loop route-to-vision-to-route transitions. Ongoing research; no public paper or code release is linked.

UAV navigationVisual groundingControl arbitrationEvidence-gated stopping

Drone Human Object Detection

ROS2 package for real-time detection and 3D localization of humans and objects using drone-mounted RealSense camera with GroundingDINO zero-shot detection.

ROS2GroundingDINORealSensePythonDeep Learning

View on GitHub

Featured Projects

AnyGoal

GoalVLM

GoalSwarm

Skyliner

Drone Human Object Detection

Ask About MoniJesu

Featured Projects

AnyGoal

GoalVLM

GoalSwarm

Skyliner

Drone Human Object Detection