Robotics Engineer · Embodied AI · Brooklyn, NY

PRABAL KHARE

I build robot systems and study what happens when learned policies fail. Sensors, comms, localization, navigation, controls, operator tooling — and now, at NYU, whether a policy can recover when it breaks.

Perception

It sees before it moves.

A 2D LiDAR scan at the front of the chassis. Reflective-marker clustering isolates two- and four-legged shelf signatures out of the return.

Localization

PGV over RS-485.

Floor tags stream into a ROS 2 localization source backed by a tag map, with wheel-odometry fallback and safety conditions when tag confidence drops.

Navigation

Nav2 path following with zone management and sector safety gating. 150 point-to-point runs on a live factory floor.

Docking

Center check. Reposition. Or abort.

The pose is validated before the lift moves. If it fails, it retries — and if it fails again, it aborts safely instead of guessing.

SMR300

300 kg. Built to be reused.

I replaced the company's ROS 1 architecture with a ROS 2 Humble stack designed for a fleet, not a prototype — CANopen and CiA 402 drives, ros2_control, an AI camera, and an operator platform for engineers who don't write code.

Project MIRA · 2nd place, TAC Challenge Norway 2024

It had to survive the pressure hull.

An autonomous underwater vehicle where every electrical decision was final the moment the hull was sealed.

• Electrical architecture and power distribution inside a sealed AUV hull
• Hydrophone and sensor-control subsystems, PCB prototyping
• ROS / MAVROS / MAVLink control software

Autonomous VTOL UAV · Top 30 of 500, SIH 2023

Find the person from the air.

A vertical-takeoff search-and-rescue platform that flew its own scan pattern and picked victims out of thermal.

• Raspberry Pi + Pixhawk + MAVROS airframe with OpenCV and thermal imaging
• TensorFlow Lite and YOLOv5 inference on board
• Autonomous scan patterns and a custom battery management system

Kurat · Autonomous conversational robot

Three brains, so one can fail.

Perception, cognition and action are separated on purpose — each can degrade or recover without taking the others down.

• RealSense D435i + Jetson Orin Nano
• YOLOv8 perception, Whisper speech, ORB-SLAM3 localization, Nav2 navigation
• Real-time interaction under a modular autonomy split

RobotDrawing · ABB IRB140

I taught it to route. It lost.

Robotic drawing reframed as a stroke-level travelling salesman problem, solved by a learned router — which then failed to beat the classical solver it was meant to replace.

• GNN + Pointer Network, imitation learning then reinforcement learning
• Simulated annealing: 26.8 pen-ups at 195 ms. The learned router: 28.8 at 10.9 s
• Slower and worse, across 28 images. Measuring it properly is how I know

OpenDroneKit · Drone Entry

Find the crack. Then predict it.

Structural inspection that doesn't stop at detection — the crack gets projected onto the reconstruction and simulated forward in time.

• ~40,000 defect images through YOLOv8 detection and U-Net segmentation
• Cracks mapped onto COLMAP 3D reconstructions
• FEniCSx simulation estimates propagation and flags severity

Continuum · Open source

Git, but for an agent's memory.

Switch from Claude Code to Codex to Gemini without re-explaining the codebase, the bug, or what you already tried.

• Append-only SQLite event log → checkpoints → materialized current.md
• log / diff / blame / restore / branch / merge, plus semantic search
• Local-first — nothing leaves .continuum/ by default

Ripple · Active research

It stopped. Now what?

A site engineer that watches a Nav2 robot fail, recovers inside a budget you set, and asks a person when the evidence says the problem is physical — then remembers what worked at that spot.

• The model only emits tool calls; the edge validates, journals, then acts
• A keepout counts once the mask and costmap show it. An arrival counts once odometry settles
• 123 unit tests, recovery 13/13 in simulation — and never yet on physical hardware

SO101 · SmolVLA · Active research

Rule out the wiring first.

Most reported manipulation failures are not policy failures. They are camera ordering, normalisation or action scaling wearing a policy's name — so the first result here is that replayed actions reproduce their demonstrations.

• Leader and follower on the LeRobot SO101, two cameras, MuJoCo evaluation
• Recorded actions replayed through the evaluator reproduce the demonstrations
• No performance claim yet. That is the honest state of it