Skip to content
View 22kashaf-khan's full-sized avatar

Highlights

  • Pro

Block or report 22kashaf-khan

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
22kashaf-khan/README.md

Kashaf Khan

Robot Learning & Simulation · Isaac Lab · MuJoCo · PPO · Sim-to-Real · Berlin, Germany

LinkedIn Email


I build and validate learned controllers for humanoid robots in simulation.

Currently writing my M.Sc. thesis with Fraunhofer IEM on Digital Twin-supported reinforcement learning for a humanoid robot: URDF/USD asset pipeline, PPO training in Isaac Lab, and sim-to-real transfer. Alongside that I work as an AI Solutions Consultant at Siemens Mobility, benchmarking and prototyping LLM systems for enterprise use.

Most of my public work is about a question I find more interesting than "can it walk": how does a learned controller behave when the model, the contact parameters, or the training objective change, and how do you measure that reproducibly?


Selected Work

Project What it is Stack
Humanoid Simulation Robustness Benchmark Trained and froze a Unitree H1 PPO locomotion policy, then built an independent benchmark around it: friction, mass, and actuator-effort perturbations across 5 seeds. 95 runs, 9,500 episodes, 44 tests, CI. Includes a custom mechanical-power reward and a matched A/B study: 43.9% power reduction at unchanged survival, but worse robustness under friction and mass mismatch. Isaac Lab · RSL-RL · PPO · PyTorch · pytest · GH Actions
MuJoCo Contact & Actuator Playground A 2-link leg written from raw MJCF, used to work through contact solver tuning (solref), position vs. velocity vs. torque actuators, and sensor logging. Written up as findings, including the ones that turned out to be modelling mistakes. MuJoCo · MJCF · Python
Scholar Multimodal RAG Retrieval pipeline over scientific papers: hybrid dense/sparse retrieval, table-aware PDF ingestion, cross-encoder reranking, RAGAS evaluation. Weaviate · LangChain · FastAPI · Docker

Stack

Simulation & RL

Isaac Sim Isaac Lab MuJoCo RSL-RL PPO Gymnasium USD

ML / DL

PyTorch HuggingFace scikit-learn XGBoost

Engineering

Python Bash SQL pytest GitHub Actions Docker FastAPI MLflow

LLM Systems

LangChain Weaviate RAGAS QLoRA


Background

M.Sc. Artificial Intelligence, BTU Cottbus-Senftenberg, 2023-2026
B.Sc. Software Engineering, SSUET Karachi

Before robotics: three years of production ML. Computer vision and NLP models at Hackerspace Karachi, LLM-powered customer support at Daraz (Alibaba Group), and an AI/IoT waste-segregation startup that won the National Idea Bank 2022.


Side project: I built an AI twin you can ask about my work. Gemini, Vercel, GitHub Pages.


Open to robotics and RL engineering roles in Germany and Europe. Happy to talk about locomotion, evaluation methodology, or anything sim-to-real.

22kashaf.khan@gmail.com

Pinned Loading

  1. humanoid-simulation-robustness-benchmark humanoid-simulation-robustness-benchmark Public

    Reproducible humanoid locomotion benchmarking in Isaac Lab with physics perturbations, robustness sweeps, and multi-seed evaluation.

    Python

  2. mujoco-contact-actuator-playground mujoco-contact-actuator-playground Public

    Hand-built MuJoCo leg model exploring contact tuning, actuators, and sensors — part of a robotics simulation portfolio series

    Python

  3. Scholar-Multimodal-RAG Scholar-Multimodal-RAG Public

    Production-grade RAG pipeline for scientific papers — PDF ingestion with table extraction, SSE streaming, multi-tenancy, and Weaviate vector store

    Python

  4. Tiangong2Pro-Integration Tiangong2Pro-Integration Public

    Python