Om Mandhane
Robotics engineer — embodied AI & generalist robotics
I want to build robots that can take a language instruction and carry it out reliably in the real world. So far my work has focused on the perception and data side of that: fusing LiDAR and vision for real-time navigation, and building low-cost teleoperation tools to collect the demonstrations that VLA models are trained on. I have a B.E. in Automation & Robotics, with a minor in AI & ML, from VESIT, Mumbai, and I'm looking to continue into a research-focused master's in world models, VLA, and humanoid manipulation.
Updates
Research interests
My research sits at the intersection of reinforcement learning, world models, vision-language-action (VLA) models, and humanoid manipulation. I'm interested in generalist manipulation policies for humanoid robots that use a predictive world model to anticipate the consequences of an action before taking it, rather than only reacting.
- Reinforcement learning
- World models
- Vision-Language-Action models
- Humanoid manipulation
- Whole-body & bimanual control
- Generalist, command-driven policies
Publications
Robust 2D Traversability Mapping for Construction AMRs via Failure-Mode-Aware Fusion of LiDAR Geometry and Monocular Semantics
Accepted for presentation at an IEEE/RSJ IROS 2026 workshop
Phone2Act: A Low-Cost, Hardware-Agnostic Teleoperation System for Scalable VLA Data Collection
Accepted at IEEE ICMRE
Return-Absence Reasoning: Self-Supervised Traversability Estimation for Unstructured Construction Sites
Submitted to IEEE ICRA 2027
Experience
Aug 2026
Robotics Engineering Intern
- Evaluated and deployed monocular metric depth models (Metric3D v2, UniDepth) on an NVIDIA Jetson AGX Orin using ONNX and FP16 TensorRT, enabling real-time point cloud generation for a downstream navigation pipeline.
- Built a real-time multimodal perception stack in ROS 2 fusing 3D Livox LiDAR geometry with a custom 28-class YOLOv26 segmentation model, including a failure-mode-aware deterministic fusion engine that mitigates geometric false positives (ramps) and false negatives (wet mud).
- Built a SAM 3-assisted annotation workflow to curate a multimodal construction dataset.
Mar 2025
Project Intern
- Developed a ROS 2 software stack for a custom indoor mobile robot, including URDF design, Gazebo simulation, and real-time hardware interfacing.
- Integrated and tuned the Nav2 stack for autonomous navigation in simulation; implemented and optimized VFH for dynamic obstacle avoidance in simulation and on the real robot.
Jun 2025
Technical Instructor
- Led two 3-day workshops on ROS 2 and Gazebo for undergraduates, run during winter and summer breaks in collaboration with ISA VESIT.
Jul 2024
Summer Intern
- Built a ROS 2 UR5 manipulation pipeline in Gazebo using the TRAC-IK solver, with a custom C++ pose-input interface for command-line end-effector control.
Projects
Mar 2026
Phone2Act — hardware-agnostic teleoperation for VLA data collection
A low-cost 6-DoF teleoperation framework that turns a commodity smartphone into a spatial controller via ARCore, built to address the demonstration-data bottleneck in vision-language-action robotics.
- Modular three-layer ROS 2 architecture (interface, planner, hardware bridge) decoupled through standardized topics, with workspace clamping and a zero-jump safety filter for real-robot deployment.
- Universal Recorder synchronizing multi-camera streams with 20Hz robot state, exporting directly to LeRobot dataset format (MP4/Parquet) — a turnkey pipeline from teleoperation to training.
- Validated on single-arm (Dobot CR5) and bimanual (LeRobot SO-101) platforms at 395ms end-to-end latency; collected 130 episodes to fine-tune NVIDIA's GR00T-N1.5-3B diffusion policy with 10-step temporal action chunking, deployed via a cloud-based inference bridge.
May 2026
Autonomous navigation with model predictive control
A real-time ROS 2 local planner solving a receding-horizon optimization problem with CasADi/IPOPT under unicycle kinematic constraints, so every planned trajectory is physically executable.
- Dynamic obstacle avoidance via an inverse-barrier cost function over downsampled 2D LiDAR scans mapped into repulsive potential fields.
- Sparse waypoints pre-processed into C²-continuous reference paths with SciPy cubic B-splines, plus a monotonic progress tracker to prevent path-jumping on sharp hairpin maneuvers.
- A cruise-control cost function to penalize deceleration and keep forward progress, with solver warm-starting inertia to eliminate saddle-point oscillations.
Mar 2025
HexBot — SLAM & navigation
A ROS 2 differential-drive robot using SLAM Toolbox for real-time mapping and Nav2 for path planning, with a VR-based streaming interface for remote monitoring.
Feb 2025
Warehouse drone
Built for the IIT Bombay e-Yantra Robotics Competition: an autonomous drone system in a simulated warehouse using ROS 2 and Gazebo, with RRT* for dynamic path planning and computer vision for package recognition.
- Tested Threshold-Based Gain Scheduling and integral windup protection for PID stability and accurate stopping in front of packages.
Feb 2024
GeoGuide
An autonomous three-wheeled robot for the IIT Bombay e-Yantra Robotics Competition, guided by an overhead camera system.
- AlexNet-based transfer learning for disaster-scene image classification, with Dijkstra's algorithm for priority-based navigation between event locations.
- ESP32 Wi-Fi communication and QGIS-based localization.
Honors & awards
- 2026Cleared GATE 2026 — Computer Science, Data Science & AI
- 2025ISA General Education Endowment Scholarship ($2,000), for outstanding potential in automation
- 2025Qualified Prototype Stage (Stage 3), Vishwakarma Awards 2025 — Smart Mobility
- 2024–25Tecnimont (MAIRE) Scholarship for academic excellence
- 2023–25Stage 2 Qualifier, IIT Bombay e-Yantra Robotics Competition (2023–24 and 2024–25)
- 20247th place, India Automation Games