Mitanshu Goel
Robotics & AI Engineer
I build the pipeline that turns human demonstrations into robot policies. Teleop rigs, synchronised capture, and the π0.5 fine-tune at the end of it. Most recently Physical AI at Nferent AI — my third robotics internship, after SarthakAI and NextUp Robotics. Open to Physical-AI and robotics roles.
See the work ↓ Read the CV →Every clip is real hardware, apart from the two labelled as simulation. Press play, or filter by where the work was done.
Where each one was built: three internships, a student robotics lab, and my own time.
Two industrial arms copy my hands live through a Meta Quest 3 that never goes on my face — it hangs on my chest as a tracking base, so the operator wears nothing. I wrote the real-time C++ loop behind it: Cartesian servoing at 125 Hz, pose in over UDP, bad tracking frames dropped rather than passed on.
A five-finger, 20-motor hand I programmed at Nferent AI. A camera reads your gesture and the hand throws its own move back.
I drove a Franka FR3 and an Elite CS66 by hand to record a 10-task LeRobot dataset, then fine-tuned a π0.5 vision-language-action policy on it and worked on its inference path.
A small humanoid I built at SarthakAI. It spots objects with YOLOv8 and answers questions out loud, 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, validated them in simulation, then fixed the URDF mismatches that were breaking trajectories on the real arm.
I told one simulated robot dog that a shove was coming and left an identical one in the dark, then shoved both about 148,000 times. The warned one falls 5× less often at 250 N. Take the warning away and it falls nine times more often than the robot that never had one.
An 18-DOF walker built with the A.T.O.M. robotics lab at MAIT. Mine is the control node and the closed-form leg IK that replaced a hand-tuned angle table. It runs on the real robot; the clip here is Gazebo.