Mitanshu Goel
Robotics & AI Engineer
Most recently, Physical AI at Nferent AI — 6-DOF arms, dexterous hands, and the teleoperation & imitation-learning pipelines that turn human demonstrations into robot policies.
See the work ↓Press play on anything — every clip is short. Filter by where it was built.
Where each project was built — three internships and my own time.
Two industrial arms copying my hands live through a Meta Quest 3, built at Nferent AI on a real-time C++ control stack I wrote — with a clutch so I can re-grip mid-task without the robots wandering.
A five-finger, 20-motor hand I programmed at Nferent AI — it reads your gesture through a camera and throws its own move back.
Driving a Franka research arm at Nferent AI behind a safety stack I wrote, recording a 51-episode manipulation dataset — 2.1 hours of RGB-D — in LeRobot format.
The capture-and-sync software I built at Nferent AI — two motion-capture gloves and three depth cameras held to one clock, under 15 ms of drift at the 95th percentile, across 45 episodes.
A small humanoid I built at SarthakAI — it spots objects with YOLOv8 and answers questions by voice, gated by a wake word.
My first internship, at NextUp Robotics — I wrote the MoveIt interfaces (joint, Cartesian-pose, and waypoint motion) with KDL inverse kinematics for a 6-DOF arm, checked in simulation, then run on the real hardware.
An 18-DoF hexapod walker — the inverse-kinematics gait engine that turns a foot target into joint angles is mine. Walking here in Gazebo simulation.
A 51-million-parameter GPT — better perplexity than GPT-2 on the same data (16.85 vs 24.68). Plus LoRA and QLoRA fine-tunes with hand-rolled distributed training; the loss curve above is the Mistral 7B Reddit run, plotted from the repo's trainer_state.json.