model-based-control

An open-source systems and controls toolbox for Python3

Fine-grained robot control with Jev, physics previews, and configurable LIBERO tasks.

Height Control and Optimal Torque Planning for Jumping with Wheeled-Bipedal Robots

JAX compilation of RDDL description files, and a differentiable planner in JAX.

PyTorch implementation of "Learning Stable Deep Dynamics Models" (https://papers.nips.cc/paper/9292-learning-stable-deep-dynamics-models), with extensions to controlled dynamical systems.

Official python code of "Guaranteed Robust Nonlinear MPC via Disturbance Feedback", ICRA 2026

We compare model-free and model-based methods in the context of battery control.

[NeurIPS 2024] Input-to-State Stable Coupled Oscillator Networks for Closed-form Model-based Control in Latent Space

Wrappers for reinforcement learning algorithms (i.e. stable baselines 3, RLlib) to work with pyRDDLGym.

Gurobi compilation of RDDL description files to mixed-integer programs, and optimization tools.

Model based control examples for legged robots

For the consolidation of my personal model predictive control (MPC) library. Example cases also given.

Phone videos of my own tabletop pushes train a small world model that plans pushes for a simulated Franka Panda in PyBullet.

Sketches of model-based control ideas for gym classic control

Code and data for the preprint "Smooth Surrogate Models Approve Impossible Firing Patterns." A smooth surrogate of a Hodgkin-Huxley f-I curve silently certifies firing rates the real neuron cannot produce (the Type-II gap and depolarization block); a feasibility screen against the true f-I removes the failure.

AI-assisted MinAtar Breakout experiment: directly construct neural-network weights from rules and dynamic programming, without training the target network. Includes three iterations with development feedback, checkpoints, raw results, reproducible comparisons, and lessons learned. Feedback and improvements welcome.

Official release candidate for Dynamic-Neighbourhood Quality-Guided World-Model Planning (DNQ-WM)

PushT world-model planning with a frozen dynamics model: look ahead to choose actions, then re-anchor and replan when prediction and reality diverge.

Safe adaptive residual reinforcement learning for robotic manipulation under model mismatch, combining physics-based control, Soft Actor-Critic, uncertainty-aware residual dynamics, and hard HOCBF safety.