Sim-to-Real
Robotic Throwing in Isaac Lab
Video coming soon - research internship., summer 2026.
Robotics Portfolio
Kishan Grewal
Robotics undergraduate building robot learning, control, embedded, and autonomy systems.
Karlsruhe Inst. of Technology
Robot Learning
Extend Robotics
Vision-Language-Action
Policy Deployment
Franka Panda VLA Deployment
Deployed SmolVLA, ACT, and Pi0 visuomotor policies in closed loop on a Franka Panda for contact-rich industrial socket insertion, fine-tuned on demonstrations collected through a Meta Quest XR teleoperation pipeline I built end-to-end.
University College London
Control & Embedded
A cart-pole balancing robot built from scratch - laser-cut chassis, Arduino Giga R1, IMU and encoder sensing, and a real-time state-feedback control stack written and tuned on hardware.
Disturbance Rejection
Tap-Test Recovery
LQR state-feedback driven by a Kalman estimator of the full cart-pole state, rejecting an impulse applied to the pendulum tip and returning to vertical without overshooting the cart limits.
Large-Angle Recovery
Recovery from 20-degree Tilt
Same controller catching the pendulum from a 20-degree initial offset - well outside the small-angle linearisation the model is derived around - and settling to upright.
Tracking Under Motion
2 m Autonomous Traverse
Cascaded control - outer state-feedback loop computing desired force, inner PID controlling wheel speed - carrying the balanced pendulum a controlled 2 metres across the floor without falling.
MangDang
Quadruped Autonomy
ROS 2 packages for a Mini Pupper 2 fleet, built during a summer internship at MangDang's Hong Kong office via the HKSTP Global Internship Programme.
Vision-in-the-Loop
Person Tracking
ROS 2 tracking pipeline fusing YOLO detections with IMU readings through a PID controller, holding yaw and pitch alignment to the target within three degrees during follow behaviour.
Autonomous Navigation
SLAM and Navigation
Indoor autonomy stack deployed on hardware - SLAM Toolbox for mapping, Nav2 for planning, AMCL for localisation - with planners and costmaps tuned for the noisy odometry that comes with quadruped locomotion.
Multi-Robot Coordination
Four-Robot Fleet
C++ coordination layer for four quadrupeds, with EKF state estimation and closed-loop heading control correcting the yaw drift that builds up under gait - enough that the robots stay in alignment without external reference.