forward pass · inference mode

Reza Kakooee

Reza Kakooee

Applied AI Researcher @FHNW, Previously PhD ETH Zurich

Zurich, Switzerland

I build production AI across Robotics, RL, LLMs, VLAs, World Models, 3D-reconstruction, computer vision, and speech.

forward() → GitHub LinkedIn X Scholar Email
L1

Research Focus

latent_space(domains)
robotics reinforcement-learning large-language-models vision-language-action computer-vision speech-synthesis
L2

Experience & Education

residual_path(trajectory)

Applied Machine Learning Researcher

FHNW, Switzerland · 2024 → present

RL, LLMs, VLA, CV, speech. Leading Swiss German TTS/STT and 3D reconstruction.

PhD in AI-Aided Design

ETH Zurich · 2019 → 2024

Deep RL for Architectural Space Layout Design. Built SpaceLayoutGym.

Invited Lecturer

Constructor Learning Institute, Zurich · 2022 → present

ML, Deep Learning, NLP, Computer Vision.

Applied ML Researcher

Hochschule Luzern (HSLU) · 2019 → 2022

ML for real-world NLP, CV, time series.

Research Assistant — Decision Neuroscience

ETH Zurich · 2018 → 2019

Human decision-making via drift-diffusion models.

MSc, Electrical & Control Engineering

Ferdowsi University of Mashhad · 2009 → 2012

Adaptive RL for offering prices in electricity markets.

BSc, Electrical Engineering

Shahrood University of Technology · 2004 → 2009

Artificial neural networks, control systems, signals & dynamic systems.

L3

Featured Works

dense_features · top_k
Multi-Agent RL2026

Can multiple AI agents solve the piano-movers puzzle together?

Five and ten RL agents, each seeing only its own small view, learn to carry a T-shaped load through two narrow slits with no leader and no messages.

Project page →
Multi-Agent RL2026

Does a leader emerge when AI agents solve the piano-movers puzzle together?

Changing which agents know the goal reveals when a leader emerges, and when a swarm remains genuinely leaderless.

Project page →
RL2026

Can an AI agent solve the piano-movers puzzle that ants solve?

An RL agent learns the piano-movers maneuver that ants solve collectively.

Project page →
Speech · FHNW2026

Text vs. Phoneme Intermediates for Swiss German TTS

SwissText 2026 paper comparing text and phoneme intermediates for low-resource Swiss German text-to-speech. DE-TTS, CH-TTS, and PH-TTS pipelines.

Project page →
Humanoid · VLA2026

AgiBot G2 Policy Lab: Humanoid Manipulation & Navigation

Adapting a 3B Vision-Language-Action model (π0.5) to the AgiBot Genie G2 humanoid: MuJoCo and Isaac Sim, 10,000 expert demonstrations, and PPO that raises success from 32.6% to 67.5%.

Project page →
Navigation · RL2026

PhyNav-1: Learning to Navigate, Failing to Recover

Teaching a simulated robot ObjectNav across 10,000 procedural houses: privileged expert imitation, PPO fine-tuning, and test-time heuristics.

Project page →
Robotics · MuJoCo · RL2026

MicroDuck Grows Up: Eight Control Experiments

Eight MuJoCo simulation experiments with one little robot: a wobble board, painting, crawling, a swing, a narrow beam, a human bridge, a rope traverse, and a trampoline flip.

Project page →
Robotics · MuJoCo2026

Can a MicroDuck Be a Bridge for Its Sister?

Two MicroDuck robots attempt a cooperative bridge crossing: one lies across a gap while the other walks over it.

Project page →
Robotics · VLA · RL2026

Teaching a UR5e to Pick and Place with VLA & RL

From broken dense rewards and 0% RL to a 50/50 Pi0.5 policy: fine-tuning an open-weights 3.5B VLA with online PPO in MuJoCo from camera pixels and language.

Project page →
Robotics2026

Self-Refining Agentic Supervisor

Robot deployment copilot coordinating simulation, validation, failure discovery, and retraining. RL-based decision framework for deployment readiness.

RL · ETH Zurich2025

Combining Behavior Cloning and RL for Space Layout Design

JCDE 2025 paper. Evaluates BC+PPO agents in SpaceLayoutGym for architectural spatial planning.

Project page →
Speech · FHNW2024

Swiss German TTS & STT

State-of-the-art Text-to-Speech & Speech-to-Text for Swiss German dialects. WER/CER & MOS evaluation pipelines, phoneme-based TTS, open-sourced parliamentary speech corpus.

stt4sg.fhnw.ch →
NLP · FHNW2024

High German → Swiss German Translation

Translation models across multiple Swiss German dialects. Parallel data collection, tokenization, and benchmarking with automatic metrics.

Language Learning2026

Reza's Deutsch Hub

A personal German study hub for grammar, vocabulary, and structured daily practice.

Open hub →
Language Learning2026

Deutsch A1 Studio

Interactive A1 learning notebooks for German grammar, sentence building, cases, vocabulary, and listening practice.

Open studio →
Computer Vision · FHNW2024

3D Reconstruction from 2D

Fine-tuned 3D reconstruction on proprietary data. Co-authored a 1M CHF funding proposal. Deployed at identic.ai.

identic.ai →
Diffusion2024

DreamBooth · SDXL

Fine-tuned Stable Diffusion-XL with LoRA. Live photo booth at Enter.ch Museum.
RL · ETH · Open Source2024

SpaceLayoutGym

Open-source space-layout simulator for RL training. Core of PhD research on deep RL for spatial planning.
RL · ETH Zurich2023

Reimagining Space Layout Design through Deep RL

JCDE 2023 paper introducing SpaceLayoutGym, laser-wall partitioning, and PPO agents for generative space layout design.

Project page →
Robotics · Education2026

A Robot That Follows an Instruction

An interactive book explaining how robot policies turn observations and language instructions into physical actions.

Read the book →
L4

Publications

peer_reviewed · attention()
  1. Kakooee, R., et al. — Text vs. Phoneme Intermediates for Low-Resource Swiss German Text-to-Speech. SwissText, 2026 Project page →
  2. Kakooee, R., Dillenburger, B. — Enhancing Architectural Space Layout Design By Pretraining Deep RL Agents. J. Computational Design & Engineering, 2025 Project page →
  3. Kakooee, R., Dillenburger, B. — Illuminating Spaces: Deep RL and Laser-Wall Partitioning. 27th Generative Art Conf., 2024
  4. Kakooee, R., Dillenburger, B. — Reimagining space layout design through deep RL. J. Computational Design & Engineering 11.3, 2024 Project page →
  5. Timmel, V., Paonessa, C., Kakooee, R., et al. — Fine-tuning Whisper on Low-Resource Languages for Real-World Applications. arXiv:2412.15726, 2024
  6. Kakooee, R., Dillenburger, B. — Design Process is a Reinforcement Learning Problem. Deep RL Workshop, NeurIPS 2022
  7. Kakooee, R., Dillenburger, B. — RLDesigner: Spatial Layout Planning as an MDP. 15th European Workshop on RL, 2022
  8. Bernhard, M., Kakooee, R., et al. — TopoGAN: Topology Optimization with GANs. Advances in Architectural Geometry, 2021
  9. Akizuki, Y., Bernhard, M., Kakooee, R., et al. — Generative Modelling with Design Constraints. CAADRIA, 2020
  10. Yousefi, A., Kakooee, R., et al. — Predicting Learning Dynamics via Variational Bayes. IEEE EMBC, 2017
L5

Project Archive

batch_size = 24
  1. Swiss German TTS & STT FHNW · 2024 → present · Speech synthesis & recognition
  2. Vision-Language-Action Models for Robot Control FHNW · 2024 → present · Robotics
  3. Self-Refining Agentic Supervisor for Robot Deployment FHNW · 2024 → present · Robotics
  4. High German → Swiss German Translation FHNW · 2024 → present · NLP
  5. 3D Reconstruction from 2D Images FHNW · 2024 → present · Computer Vision
  6. NLP for Risk Management & Safety Analysis FHNW · 2024 → present · NLP
  7. Patent Similarity Detection for Freedom-to-Operate FHNW · 2024 → present · LLM · RAG
  8. DreamBooth Personalization for Stable Diffusion FHNW · Diffusion · LoRA
  9. SpaceLayoutGym ETH Zurich · 2019–24 · RL · Open Source
  10. Multi-Task Learning for Segmentation, Depth & LiDAR ETH Zurich · Course Project
  11. LSTM-Based Forecasting & Clustering for Seismic Data ETH Zurich · Voluntary Collaboration
  12. COVID-19 Case Rates Forecasting & Chest X-Ray Clustering ETH Zurich · Voluntary Collaboration
  13. Drift-Diffusion Model Parameter Estimation ETH Zurich · 2018–19 · Decision Neuroscience
  14. Medical Language Models for Semantic Similarity HSLU · 2019–22 · NLP
  15. Sentiment Analysis on German Product Reviews HSLU · 2019–22 · NLP
  16. NSFW Image Detection & Zero-Shot Classification HSLU · 2019–22 · Computer Vision
  17. Hidden Markov Model for Multi-Crypto Price Forecasting HSLU · 2019–22 · Time Series
  18. Dynamic Pricing for Hotels HSLU · 2019–22 · ML
  19. Anomaly Detection for Sensor Measurement Errors HSLU · 2019–22 · Time Series
  20. CNN Classifier for Selfie Detection HSLU · 2019–22 · Computer Vision
  21. Logo Similarity with VGG16 & Inception HSLU · 2019–22 · Computer Vision
  22. Customer Loyalty & Churn Prediction S-Bank · 2017 · ML
  23. Adaptive MPC for Pick-and-Place Robot Arms TMU · 2016 · Control
  24. Fuzzy Controller for 3-Degree Parallel Robot FUM · 2011 · Control
L6

Foundations

inductive_bias(ctrl + math)

Control Engineering

  • Fuzzy Control — 3-degree parallel robot controller design & tuning
  • Model Predictive Control — Adaptive MPC for pick-and-place robot arms
  • Robust & Multivariate Control — H∞, LQG, and MIMO controller design for uncertain systems
  • Signal Processing — Sensor fusion, filtering, and anomaly detection pipelines

Mathematics

  • Optimization — Convex & non-convex, gradient methods, constrained optimization
  • Dynamical Systems — Phase portraits, stability analysis, Lyapunov methods
  • Probability & Statistics — Bayesian inference, stochastic processes, MCMC
  • Decision Theory — Drift-diffusion models, evidence accumulation, bounded rationality
L7

Stack

model.parameters()

Languages

Python · MATLAB · C++ · JavaScript · SQL · R

ML / Deep Learning

PyTorch · TensorFlow · Keras · scikit-learn · Avalanche

NLP & LLMs

HuggingFace · LangChain · TRL · Accelerate · Diffusers

Reinforcement Learning

RLlib · TorchRL · Stable-Baselines · Isaac Sim · Isaac Lab · MuJoCo · Gym

L8

Writings & Insights

decoder(thoughts)
  1. Humanoid faces should make intent legible LinkedIn · Aug 2026
  2. When an RL navigation agent walks backward LinkedIn · Aug 2026
  3. Teaching an AI agent the ant piano-movers maneuver LinkedIn · Aug 2026
  4. MicroDuck learns dynamic balance X · Sep 2026
  5. MicroDuck Picasso LinkedIn & X · Sep 2026
  6. MicroDuck learns to crawl backward LinkedIn & X · Sep 2026
  7. MicroDuck on a playground swing LinkedIn · Sep 2026
  8. Can an AI-agent swarm solve the piano-movers puzzle? LinkedIn · Sep 2026
  9. Can a MicroDuck be a bridge for its sister? LinkedIn · Sep 2026
  10. Intent-to-Motion: Control Stack Widget Interactive · Jul 2026
  11. Choosing the right intermediate representation for Swiss German TTS LinkedIn · Jul 2026
  12. Why validation loss can rise while accuracy improves LinkedIn · Aug 2026
  13. What the AI Community Currently Lacks LinkedIn · Sep 2024
  14. AI is My Last Hope LinkedIn · Sep 2024
  15. Teaching is All About Knowledge Compression LinkedIn · Sep 2024
  16. My Journey with AI LinkedIn · Sep 2024
  17. AI & LLMs LinkedIn · Jul 2024
  18. Reinforcement Learning & Deep Learning LinkedIn · Jul 2022
  19. Design Process is a Reinforcement Learning Problem LinkedIn · 2022
  20. Deep RL for Architectural Space Layout Design — Part III Medium · 2022
  21. Deep RL for Architectural Space Layout Design — Part II Medium · 2022
  22. Deep RL for Architectural Space Layout Design — Part I Medium · 2022
  23. Deep Double Descent LinkedIn · Dec 2019
L9

Output

softmax() → p(collab)
Get in touch

Open to research collaborations, speaking invitations, and industry projects.

GitHub LinkedIn X Scholar Email

© 2025 Reza Kakooee · loss converged