Projects

Things I've built, published, and open-sourced.

Engineering projects, research prototypes, open-source libraries, and experiments that outgrew the weekend they started in. Code is public wherever I can share it, and pull requests are welcome.

01

Robotics · RL · Simulation

Helios — 6-DOF Pick-and-Place Controller

A reinforcement learning policy for a UR5e collaborative robot arm, trained end-to-end in IsaacGym and transferred to hardware with minimal fine-tuning. The core contribution is a domain randomisation schedule that gradually narrows the simulation-reality gap over training, combined with a learned force/torque residual that compensates for unmodelled contact dynamics. Achieves 94% grasp success rate across 30+ unseen object geometries without any geometry-specific calibration.

PyTorchIsaacGymROS 2Domain RandomisationUR5eForce/Torque Sensing

View on GitHub Read the paper

02

Software · Motion Planning · Open Source

Kepler — Open-Source Trajectory Planner

A modular C++ motion planning library for mobile robots and manipulators. Implements RRT*, informed RRT*, CHOMP, and TrajOpt as interchangeable planners behind a clean, unified API. Ships with Python bindings, ROS 2 integration, a comprehensive benchmark suite, and documentation that actually explains the assumptions. Now used by three independent research groups and cited in 12 peer-reviewed papers. Actively maintained with monthly releases.

C++17Python BindingsROS 2RRT* / CHOMPpybind11CMake

emreari/kepler Documentation

03

Research · Legged Robotics · ETH Collaboration

Mira — Quadruped Gait Policy

A sample-efficient reinforcement learning framework for quadruped locomotion over unstructured terrain, developed in collaboration with ETH Zürich's Robotics Systems Lab. The key contribution is a hierarchical policy architecture that separates gait generation from terrain adaptation — enabling fast adaptation to novel surfaces without retraining the base policy. Validated on the ANYmal-C platform across outdoor terrain including gravel, grass, and uneven concrete. Published at IEEE ICRA 2024.

PyTorchANYmal-CHierarchical RLRaiSimTerrain Adaptation

Read the paper Code (partial)

Modular motion planning library for mobile robots and manipulators. RRT*, CHOMP, TrajOpt with a unified API, Python bindings, and full ROS 2 integration. Used by 3 research groups.

C++

847

142

Standardised benchmarking suite for sim-to-real transfer in robotic manipulation. 12 task environments, unified metrics, and baseline results for 6 state-of-the-art algorithms.

Python

312

58

Opinionated toolkit for deploying PyTorch models to NVIDIA Jetson and Raspberry Pi hardware. Handles TensorRT conversion, latency profiling, and memory budgeting.

Python

194

41

Production-ready Docker development environment for ROS 2. GPU passthrough, VS Code devcontainer config, X11 forwarding, and automated setup scripts. Companion to the blog tutorial.

Shell

428

96

View all repositories on GitHub

Active

World Model-Based Planning for Manipulation

Exploring DreamerV3 as a backbone for dexterous manipulation planning. Current question: does the world model actually learn useful contact dynamics, or just correlations?

Active

Uncertainty-Aware Edge Deployment

Investigating Bayesian neural networks and ensembles for uncertainty quantification on NVIDIA Jetson hardware. Target: calibrated confidence with under 5ms overhead.

Active

Generative Terrain Augmentation for RL

Using a diffusion-based terrain generator to produce training environments that are structurally different from those seen during policy training — targeted curriculum rather than random.

Completed

LiDAR-Camera Calibration Automation

Automated extrinsic calibration pipeline for LiDAR-camera pairs using a learned target-free approach. Reached 0.3° angular accuracy. Released as part of rl-transfer-bench.

Completed

Modular Synthesizer MIDI Controller

A custom Raspberry Pi-based MIDI controller for my modular synthesizer setup. C++ firmware, 32 knobs, 16 buttons. Finally finished after three years. It works.

Completed

MLflow Drift Detection Plugin

A plugin for MLflow that adds statistical drift detection to logged models. Integrates with the Orion monitoring work. Currently under review for integration into the main MLflow project.

Want to collaborate on a project?

If you're working on something in robotics, machine learning, or open-source tooling, I'd like to hear about it.

Get in touch See my research

Things I've built, published, and open-sourced.

Projects with a story worth telling

Helios — 6-DOF Pick-and-Place Controller

Kepler — Open-Source Trajectory Planner

Mira — Quadruped Gait Policy

Code I've released publicly

Things I'm exploring right now

World Model-Based Planning for Manipulation

Uncertainty-Aware Edge Deployment

Generative Terrain Augmentation for RL

LiDAR-Camera Calibration Automation

Modular Synthesizer MIDI Controller

MLflow Drift Detection Plugin

Want to collaborate on a project?