Fast Online Adaptive Neural MPC via Meta-Learning

This is the official implementation of the paper:
"Fast Online Adaptive Neural MPC via Meta-Learning"
by Yu Mei, Xinyu Zhou, Shuyang Yu, Vaibhav Srivastava, and Xiaobo Tan.

---

📑 Table of Contents

• Fast Online Adaptive Neural MPC via Meta-Learning
  • 📑 Table of Contents
  • 🎥 Demonstration Video
  • 🛠️ Installation Instructions
    • Option A: CUDA 12.x GPUs (e.g., RTX A2000)
    • Option B: Blackwell GPUs (CUDA 13.x, e.g., RTX 5070)
  • 🚀 How to Run
    • 1. Van-der-Pol Oscillator
    • 2. CartPole
    • 3. 2D Quadrotor Stabilization and Tracking
      • Stabilization Results
      • Tracking Results
  • 📚 Project Structure
  • 📝 Citation

---

🎥 Demonstration Video

Watch our YouTube video showcasing the control performance on the CartPole and 2D Quadrotor environments using the proposed Fast Online Meta-MPC framework.

---

🛠️ Installation Instructions

1. Clone the repository (common to both tracks)

git clone https://github.com/yu-mei/MetaResidual-MPC.git
cd MetaResidual-MPC

2. Pick the track that matches your GPU, then follow one of the two options below:

Your GPU	Track
Pre-Blackwell cards — RTX 20/30/40 series, A-series (e.g., RTX A2000)	Option A
Blackwell cards — RTX 50 series (e.g., RTX 5070), compute capability sm_120	Option B

---

Option A: CUDA 12.x GPUs (e.g., RTX A2000)

A1. Create a conda environment

conda env create -f environment.yml
conda activate l4control

A2. Install l4casadi

Install the latest version using pip with --no-build-isolation (GPU/CUDA supported):

pip install l4casadi --no-build-isolation

🔗 Source: github.com/Tim-Salzmann/l4casadi

A3. Install acados and the acados Python interface

A3.1 Clone and build Acados

Follow the official Acados installation guide.

A3.2 Install the Acados Python interface

Follow the Python interface installation guide.

A4. Install safe-control-gym

Follow the official safe-control-gym installation guide.

A5. Override PyTorch installation

Due to version conflicts between l4casadi and safe-control-gym, it is necessary to override PyTorch:

conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.4 -c pytorch -c nvidia

⚠️ This ensures compatibility with both l4casadi and safe-control-gym.
🔧 Make sure your CUDA drivers are compatible with CUDA 12.4.

A6. Fix installation issues (if any)

If you encounter any remaining errors, manually install the missing or incompatible packages.
Package versions may vary depending on your system environment.

---

Option B: Blackwell GPUs (CUDA 13.x, e.g., RTX 5070)

⚠️ Why a separate track? Blackwell cards (RTX 50 series) have compute capability sm_120, which PyTorch only supports from 2.7+ with CUDA ≥ 12.8 builds. The Option A install (pytorch==2.5.1 + pytorch-cuda=12.4) imports fine but fails at the first GPU op with CUDA capability sm_120 is not compatible with the current PyTorch installation. In addition, the -c pytorch conda channel no longer publishes new versions, so PyTorch is installed via pip wheels here, and l4casadi moves to the last step so that it compiles against the final PyTorch.

💡 The CUDA Version: 13.x shown by nvidia-smi is the driver's maximum supported runtime, not an installed toolkit. The pip wheels bundle their own CUDA runtime, so no system CUDA toolkit is required for PyTorch.

B1. Create a conda environment

conda env create -f environment.yml
conda activate l4control

B2. Install acados and the acados Python interface (same as A3 — unchanged)

The acados C library is CPU-only and unaffected by the GPU swap.

B2.1 Clone and build Acados

Follow the official Acados installation guide.

B2.2 Install the Acados Python interface

Follow the Python interface installation guide.

B3. Install safe-control-gym (same as A4)

Follow the official safe-control-gym installation guide.

💡 Whatever PyTorch version it pulls in will be replaced in the next step — ignore it for now.

B4. Install PyTorch (Blackwell build) (replaces A5 — do not run the old conda command)

pip uninstall -y torch torchvision torchaudio triton
pip install "torch==2.9.*" "torchvision==0.24.*" "torchaudio==2.9.*" \
    --index-url https://download.pytorch.org/whl/cu130 --no-cache-dir

💡 torch ≥ 2.7 is the hard minimum for sm_120; the 2.9 series is the first with CUDA 13.0 wheels, still supports Python 3.10, and is a modest API jump from 2.5.1.

Conservative fallback if the code misbehaves under 2.9 (cu128 wheels remain hosted and run fine on 13.x drivers):

pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 \
    --index-url https://download.pytorch.org/whl/cu128 --no-cache-dir

🔗 Source: pytorch.org/get-started/locally · pinned versions: pytorch.org/get-started/previous-versions

B5. Install l4casadi — last, after PyTorch (moved from A2)

pip install l4casadi --no-build-isolation --no-cache-dir

⚠️ Order matters: --no-build-isolation makes l4casadi compile and link against the currently installed PyTorch, so it must come after B4. If you ever change the torch version later, force a rebuild: pip install l4casadi --no-build-isolation --force-reinstall --no-cache-dir
🔧 Requires GCC ≥ 10. If the GPU build is not detected automatically and a system toolkit exists: CUDACXX=/usr/local/cuda/bin/nvcc pip install l4casadi --no-build-isolation.
🔗 Source: github.com/Tim-Salzmann/l4casadi

B6. Verify the GPU stack

import torch
print(torch.__version__, torch.version.cuda)   # e.g. 2.9.x  13.0
print(torch.cuda.get_device_name(0))           # NVIDIA GeForce RTX 5070
print(torch.cuda.get_device_capability(0))     # (12, 0)
x = torch.randn(1024, 1024, device="cuda")
print((x @ x).sum())                           # real kernel launch = the actual test

⚠️ torch.cuda.is_available() only checks the driver — always run the matmul line. If you see sm_120 is not compatible or no kernel image is available, a pre-cu128 or CPU wheel sneaked in: redo B4 (uninstall first, keep --no-cache-dir and the explicit --index-url).

B7. Fix installation issues (if any)

If you encounter any remaining errors, manually install the missing or incompatible packages.
Package versions may vary depending on your system environment.

💡 If you reinstall safe-control-gym later, use pip install -e . --no-deps so it does not downgrade PyTorch.
💡 If your local copy of environment.yml predates the repo cleanup and still lists nvidia-*-cu11 or triton==3.1.0 under pip:, delete those lines (or pip uninstall the packages) — they are leftovers from an old CUDA 11-era PyTorch and conflict with the Blackwell wheels.

---

🚀 How to Run

1. Van-der-Pol Oscillator

We provide several scripts under VanderPol/ for running different versions of the Van der Pol system using Meta-MPC:

Script	Description
VanderPolSys_sim.py	Simulates the nominal Van der Pol system.
VanderPolSys_real.py	Simulates the real system (with mismatched dynamics).
VanderPolSys_naive.py	Runs nominal MPC to predict the trajectories.
VanderPolSys_naive_lightmlp.py	Runs nominal MPC with a lightweight learned MLP model (learn from stratch)
VanderPolSys_naive_meta.py	Runs MPC using a meta-learned model.
VanderPolSys_Collection_Meta.py	Collects offline data for training the meta-learned model, and data file vdp_meta_nominal_residual.py is under dataset
Comparsion.ipynb	Jupyter notebook comparing performance across methods.
MetaLearning/Offline_Train_Meta.py	Training the Meta MLP offline using the data file vdp_meta_nominal_residual.py

📌 Example: Run nominal MPC + Meta MLP

cd VanderPol
python VanderPolsys_naive_meta.py

After running the different methods and saving results in the results/ folder, open Comparsion.ipynb to visualize and compare the performance.

---

2. CartPole

We provide several scripts under Cartpole/ to run different MPC controllers for the CartPole system using our Meta-MPC framework:

Script	Description
cartpole_Nominal.py	Runs MPC with a nominal (physics-based) model.
cartpole_LightMLP.py	Runs MPC using a learned residual MLP trained from scratch.
cartpole_MetaMLP.py	Runs MPC using a meta-learned residual MLP with online adaptation.
cartpole_Nominal_seeds.py	Batch test across seeds using nominal model.
cartpole_LightMLP_seeds.py	Batch test using residual MLP (non-meta).
cartpole_MetaMLP_seeds.py	Batch test using meta-residual MLP model.
MetaLearning/DataCollection_Meta.py	Collects residual training data for meta-learning. The output is saved in meta_dataset_mpc/.
MetaLearning/Offline_Train_Meta.py	Trains the Meta-Residual MLP model using the collected CSV dataset.
meta_dataset_mpc/cartpole_meta_residual_mpc.csv	CSV dataset collected from DataCollection_Meta.py used for offline meta-learning.
Comparsion.ipynb	Jupyter notebook to visualize and compare results across all methods.

📌 Example: Run Meta-MPC with Online Adaptation

cd Cartpole
python cartpole_MetaMLP.py

📌 Example: Collect Residual Data for Meta-Training

python MetaLearning/DataCollection_Meta.py

📌 Example: Train Meta Residual MLP Offline

python MetaLearning/Offline_Train_Meta.py

After running all variants, results will be saved in the results/ folder.
Open Comparsion.ipynb to visualize metrics such as RMSE, trajectory tracking, and adaptation efficiency.

Left: Nominal MPC   |   Middle: Neural MPC + Residual MLP   |   Right: Neural MPC + Residual Meta-MLP

---

3. 2D Quadrotor Stabilization and Tracking

We provide two folders for 2D Quadrotor control tasks using our Meta-MPC framework:

• Quadrotor_2D_Stabilization/: for stabilization tasks
• Quadrotor_2D_Tracking/: for reference trajectory tracking

Each folder contains scripts to run, which is similar as CartPole system:

📌 Example: Run Meta-MPC for Stabilization

cd Quadrotor_2D_Stabilization
python quadrotor2D_Meta.py

📌 Example: Run Meta-MPC for Tracking

cd Quadrotor_2D_Tracking
python quadrotor2D_Meta.py

Stabilization Results

Left: Nominal MPC   |   Middle: MPC + Residual MLP   |   Right: Neural MPC + Residual Meta-MLP

---

Tracking Results

Left: Nominal MPC   |   Middle: MPC + Residual MLP   |   Right: Neural MPC + Residual Meta-MLP

---

📚 Project Structure

MetaResidual-MPC/
├── assets/                      # Demo GIFs and figures used in README (e.g., CartPole_MetaMLP.gif, Quadrotor_*.gif)
├── Cartpole/                    # Code for CartPole experiments
├── Quadrotor_2D_Stabilization/  # Code for 2D Quadrotor stabilization tasks
├── Quadrotor_2D_Tracking/       # Code for 2D Quadrotor trajectory tracking tasks
├── VanderPol/                   # Code for Van der Pol oscillator experiments
├── environment.yml              # Conda environment file
└── README.md                    # Project documentation

---

📝 Citation

If you find our work useful, please consider citing:

@article{mei2025fast,
  title={Fast Online Adaptive Neural MPC via Meta-Learning},
  author={Mei, Yu and Zhou, Xinyu and Yu, Shuyang and Srivastava, Vaibhav and Tan, Xiaobo},
  journal={IFAC-PapersOnLine},
  volume={59},
  number={30},
  pages={377--382},
  year={2025},
  publisher={Elsevier}
}

---