anmol@kcl:~ maths · robotics · liquid neural networks
click or press any key to skip
anmol@kcl:~ maths · robotics · liquid neural networks
Final-year Mathematics with Statistics student at King's College London, and a Visiting Student AI Researcher at Oxford Dynamics.
This summer I was a Robotics Intern at Oxford Dynamics. I came in with no robotics background, learned how a 6-DoF robotic arm works, built a controller for it in Python with ROS 2 and MoveIt, and recreated it in NVIDIA Isaac Sim.
Now I'm researching liquid time-constant networks as adaptive filters for my final-year project, jointly with KCL and Oxford Dynamics.
Outside research I co-lead the IBM Z Society at KCL, represent students across my faculty, and have taught Python to beginners.
I built this site myself as a tiny desktop OS, with Claude Code as my pair programmer. The source is on GitHub ↗.
Tip: open the folders on the desktop, or launch and type help.
Final-year project · KCL × Oxford Dynamics · 2026–27
Do liquid networks learn to behave like a Kalman filter?
Most neural networks process the world in fixed time steps. Liquid time-constant (LTC) networks are continuous-time models: every neuron is a small differential equation, and its time constant changes with the input it receives. A strong signal makes a neuron react quickly; a quiet one lets it integrate slowly.
A Kalman filter is the classical, optimal way to estimate the state of a system from noisy measurements. Its key quantity, the Kalman gain, decides how much to trust each new measurement against the current estimate.
An LTC neuron's input-dependent time constant plays a strikingly similar role: it decides how fast the neuron's state moves toward new input. My project investigates whether the time constants an LTC network learns actually correspond to the Kalman gain of an optimal linear filter, connecting continuous-time neural ODEs with classical state estimation.
Run as my KCL final-year project under joint academic and industry supervision, with Oxford Dynamics' research division.
In progress. Notes and results will be added to this folder as the project develops.
Projects / Arm Controller / README.md
Drive a UR7e robot arm from the command line, with the controller I wrote in Python on ROS 2 and MoveIt during my internship at Oxford Dynamics. Here it runs in your browser and moves a simulated UR7e.
client.py, and the arm.Enter Command --> and press Enter.move z 100 to lift the tool 10 cm, then move x -150, then home to go back.pick_and_place while the arm is at home.move x 100, move y -50, move z 80: move the tool along an axis, in millimetres.move rx 30, move ry -20, move rz 45: rotate the tool about its own axes, in degrees.Linear moves are measured from the robot's base unless you switch to the tool frame (see modes). In the base frame, x points forward from the robot, y to its left and z up.
mode takes two letters. The line above every prompt tells you which mode you're in.
g (global) measures moves in the base frame, l (local) in the tool's own frame.c (Cartesian) moves the tool in a straight line, j (joint space) moves all six joints together to the same target, so the tool's path may curve.mode gc is the default, and mode lj is local plus joint space.If a straight-line move can't be completed, for example because it would hit the table or pass through a singularity, the controller asks Try via joint states instead? (y/n). Answer y to reach the same target with a joint-space move.
save my_pose: save where the arm is now as a waypoint. Saved waypoints stay in your browser.my_pose: type a waypoint's name to move back to it. Built in: home, pick_above, pick, place_above, place.list_states: list every waypoint.pick_and_place: run a trajectory, a sequence of waypoints. A trajectory only starts if the arm is already at its first waypoint, so type home first if it refuses.list_trajectories: list every trajectory.help lists the commands, and help move shows one.q quits the controller. Run python3 client.py at the shell prompt to start it again, or exit to close the terminal.Invalid input/ Wrong input format and Would you like to try again? (y/n). Answer y to carry on.Two ROS 2 nodes. The client node reads commands, manages waypoints and reads the live /joint_states topic. The planner node holds the Cartesian and joint-space planners and executes the paths through MoveIt. Both run on a MultiThreadedExecutor in a background thread, so the blocking command prompt and live ROS communication can run at the same time.
In the browser there's no ROS. The terminal runs a line-by-line JavaScript port of client.py, a small IK solver and planner stand in for MoveIt, and the only collisions checked are with the table and the robot's base. The arm is the UR7e model from Universal Robots' ROS 2 description (BSD-3-Clause).
Projects / Anmol Agent / README.md
An AI version of me that lives in this site's terminal. It answers questions about my work, my studies and this website in the first person, and it runs entirely on your own GPU.
agent in any terminal on this site.you ›, like what I'm working on, what's fun to try on the site, or how this agent works. Follow-ups like "tell me more" work too.yes, and it opens on screen.exit or press Esc to leave. For a single question, use ask <question> instead.Only what's on this site: every document on the desktop, the site's README (features, things to try, terminal commands, how it was built) and my CV. It has no internet access and no memory between visits.
It's retrieval-augmented generation (RAG) running entirely in the browser:
I tested six open models through whole conversations full of trick questions. Four made up facts about me that aren't on this site, so they were dropped: Qwen2.5 0.5B, 1.5B and 3B, and Llama 3.2 1B. The two that stayed accurate are the ones you can pick: Llama 3.2 3B (1.8 GB, recommended) and Llama 3.1 8B (4.5 GB, the best answers on a strong GPU).
Built with WebLLM (Apache-2.0) and Meta's Llama models.
Open the and folders for the full story and photos.
King's College London — BSc (Hons) Mathematics with Statistics, 2024–2027
IBM Z Xplore Concepts · IBM Z & LinuxONE Community Contributor
Harwell Science Campus, Oxfordshire · defence AI & robotics
I stayed on after my internship to lead research on liquid time-constant networks as adaptive filters, run as my KCL final-year project. See the folder.
"The part I didn't expect to matter so much: the gap between code you write alone for a university project and code that ships in a real codebase."
Photos from Harwell are in the folder.
As Co-President and IBM Z Student Ambassador, I represented IBM Z and mainframe technology on campus: workshops, watch parties and a growing community of learners.
In February 2026 we ran a two-day industry speaker series with the KCL AI Society and KCL Quant Society. Day 1 brought Amazon engineers on applied generative AI, and I hosted the fireside chat. Day 2 brought an IBM director on AI and enterprise computing.
Photos from the speaker series are in the folder.
Always happy to connect and learn from more people.
Opens a message window: write your note and hit Send.
Couldn't load the CV here.
Open it in a new tab, or request it by email.