Robotics engineer · state estimation · SLAM · planning · embodied AI
I build the software that lets robots know where they are and decide where to go next.
I work across the full robotics stack, from sensor drivers and state estimation to planning, control and simulation. I've shipped it on almost every kind of robot:
| Embodiment | What I built | |
|---|---|---|
| 🚗 | Autonomous cars & pods | Took ZPod, India's first autonomous vehicle, from zero to 100k+ lines of production autonomy code and 10,000 km driven: planning, prediction, control, simulation |
| 🛵 | Two-wheelers | Navigation and motion planning; stereo + LiDAR + IMU + GPS state estimation |
| 🐕 | Quadrupeds | Autonomy and localization for legged robots |
| 🌲 | Outdoor autonomy | Off-road and outdoor navigation, robust to GPS and vision dropouts |
| 🛡️ | Defence | Defence robotics at Aeroarc: VIRA tracked UGV (15 units in active deployment), ARCV MULE quadruped (10 in active deployment), PhiSCAN 3D LiDAR scanner |
| 🔍 | Inspection & scanners | Inspection robots, scanning and mapping |
| 🦾 | Manipulators & AMRs | Bin localization (YOLOv5 + PnP), control state machines, ArUco auto-docking |
| 🌊 | Underwater: now | VIO + DVL fusion, GPU front ends, AI inspection reports, sim-trained ROV control |
Underwater is the one embodiment I hadn't done yet, so that's where I'm working now.
- State estimation & SLAM: visual-inertial, lidar-inertial, DVL/pressure fusion, factor graphs (GTSAM), fixed-lag smoothing, relocalization, sensor dropout handling
- Planning & control: path planners, trajectory optimization, MPPI, prediction, genetic-algorithm and actor-critic PID tuning, low-latency C++
- Simulation: built an agentic Unreal Engine pipeline that generates photorealistic simulators, an India-specific AV simulator for unstructured traffic, and closed-loop CI testing; Stonefish, Gazebo, MuJoCo
- GPU & real-time: CUDA perception and planning, Jetson Orin, multithreaded low-latency systems
- Embodied AI & deployment: taking learned policies and VLM-driven behaviours from simulation onto real robots (on-device inference, latency budgets, fallbacks); PPO driving policies, H-GAIL (hierarchical adversarial imitation), RL locomotion and control in MuJoCo (MJX + PPO), vision-language models for robot decision-making and inspection
| Project | What it does | Result so far |
|---|---|---|
| uw-vio-dvl | Stereo-inertial-pressure VIO with a tightly coupled DVL factor that survives dropouts | 20 s camera blackout: ATE 3.88 m → 0.66 m with DVL |
| uw-gpu-klt | CUDA KLT stereo front end | 2.0–4.9× faster, no accuracy loss |
| uw-inspect-report | Detector + VLM that writes the draft inspection report | 8/8 true findings confirmed, precision 0.89 |
| uw-calib-health | In-water stereo calibration health monitor | every ≥ 0.2° drift caught, 0 false alarms |
| uw-vslam-bench | Benchmark harness, plus the experiments that didn't work, published honestly | — |
| rov-rl-mujoco | RL for an ROV in MuJoCo (PPO in MJX): sonar-only navigation through crossing traffic + control in currents, vs VFH+ and PID | 0.4 m/s current: 2.7× lower tracking error than PID (0.048 vs 0.131 m); sonar navigation being re-validated |
| quadruped | Browser robot-dog simulator: hand-written multibody dynamics, trained trot on dynamic gravel, SPH water | trained trot, real time on flat ground |
| humanoid-mimic-rl | Booster T1 humanoid in MuJoCo MJX: PPO walking, get-up, backflip, DeepMimic + SuperMimic | walk: 0 falls in 128 × 20 s episodes, 0.47 m/s vs 0.5 commanded |
Next up: humanoid get-up, backflip and SuperMimic (training), and sonar-based RL navigation for the ROV.
| When | Where | Role |
|---|---|---|
| 2026 – now | Griffin Labs, Singapore | Autonomy Engineer: building generalisable autonomy |
| 2025 – 2026 | Aeroarc, Bengaluru | Senior Robotics Engineer: India's defence robotics; VIRA (15 in active deployment), ARCV MULE (10 in active deployment), PhiSCAN |
| 2024 – 2025 | BYTES, Bengaluru | Founding member, Robotics & AI: led tech and product; raised $500K, built the product and ran a public demo; selected for Entrepreneur First and YC Startup School |
| 2023 – 2024 | Minus Zero, Bengaluru | Head of Planning & Prediction: led ZPod's autonomy software and team from zero, from pre-seed to Series A; simulation pipeline, CI/CD, RL |
| 2022 – 2023 | Minus Zero | Founding Engineer, Autonomy: first software hire |
| 2021 – 2022 | Clutterbot (via In2peta Robotics) | Robotics Intern: bin localization, FSM control, ArUco docking, led two interns |
B.Tech ELC (robotics), Amrita Vishwa Vidyapeetham
C++ Python CUDA ROS 2 GTSAM Eigen OpenCV PyTorch JAX MuJoCo Unreal Engine NVIDIA DriveWorks Gazebo Stonefish Nav2
nehalnevle@gmail.com: always up for hard robotics problems, especially wet ones.