RAID-Net: Recurrent Attention-based for Interaction Duals Network

This repository contains the implementation of the paper "Scalable Multi-modal Model Predictive Control via Duality-based Interaction Predictions" accepted at 2024 IEEE Intelligent Vehicles Symposium

Hansung Kim ([email protected])   Siddharth Nair ([email protected])   Francesco Borrelli

Demonstration Video:

Example Simulation Results

Baseline: Full MPC with collision avoidance constraints imposed for all multi-modal predictions

Proposed: Sparse MPC with fewer collision avoidance constraints based on duality-based interaction predictions

>12x Improvement in computation time of the motion planner!

Installation

Clone repository

git clone https://github.com/MPC-Berkeley/hmpc_raidnet.git

Build environment

cd hmpc_raidnet
conda env create -n raidnet -f environment.yml
conda activate raidnet

Install dependency

pip install -e .

Installing solvers

Running the simulator requires Gurobi and IPOPT. For installing Gurobi, find the link here. For installing IPOPT (keep in mind that it comes pre-installed with Casadi),

wget -N -q "https://github.com/thomasfork/ipopt_linux/raw/main/ipopt-linux64.zip"
unzip -o -q ipopt-linux64

We provide a shell script 'grbpath.sh' which you can run in your terminal to export appropriate GUROBI related envrionment variables. Modify 'grbpath.sh' file to the paths in your machine.

source grbpath.sh

Training RAID-Net

Configuration:

The default configuration file is data/configs/params_N14.yaml. Change the root directory in the configuration files to appropriate paths in your local machine manually.

Data collection:

python scripts/data_collect.py -config_dir <path_to_parameter_config_yaml>
python scripts/data_collect.py -config_dir ./data/configs/params_N14.yaml 

Alternatively, you can download a training dataset using the default simulation environment parameters and data collection configurations (data/configs/params_N14.yaml). Modify the configuration file to change the parameters as needed.

If you choose to download and use the provided dataset, move the dataset .pkl files as shown below

├── assets
│   ├── ...
├── data
│   ├── configs
│   │   ├── ...
│   ├── expert_data
│   │   ├── <move data here>
│   ├── logs
│   │   ├── ...
│   ├── models
│   │   ├── ...
├── infrastructure
│   ├── ...
├── scripts
│   ├── ...
├── ...

Running behavior cloning:

Before training, modify the expert_trajectory_data path in ./data/configs/*yaml file that you are using. For training a RAID-Net Policy, run

python scripts/run_bc.py -config_dir <path_to_parameter_config_yaml>
python scripts/run_bc.py -config_dir ./data/configs/params_N14.yaml 

For training a Simple MLP Policy for comparison,

python scripts/run_bc_mlp.py -config_dir <path_to_parameter_config_yaml>
python scripts/run_bc_mlp.py -config_dir ./data/configs/params_N14.yaml 

(WIP) For running DAgger,

python scripts/run_dagger.py -config_dir <path_to_parameter_config_yaml>
python scripts/run_dagger.py -config_dir ./data/configs/params_N14.yaml 

Alternatively, you can download the pre-trained RAID-Net for N=14 and Num TV = 3 in our custom traffic intersection simulation environment here. After downloading the model, move the model files .pt as shown below

├── assets
│   ├── ...
├── data
│   ├── configs
│   │   ├── ...
│   ├── expert_data
│   │   ├── ...
│   ├── logs
│   │   ├── ...
│   ├── models
│   │   ├── <move models here>
├── infrastructure
│   ├── ...
├── scripts
│   ├── ...
├── ...

Evaluation

Evaluation parameters can be modified in the ./data/configs/*.yaml files. Specifically, you must change the root_dir variable in the config files to an appropriate path in your machine. Also, you can set the planner horizon length N, model names, and other training and environmental parameters as needed.

python scripts/closedloop_eval.py -config_dir <path_to_parameter_config_yaml>
python scripts/closedloop_eval.py -config_dir ./data/configs/params_N14.yaml 

The evaluation results will be recorded in ./data/models/evaluation/ including mp4 files of the closed-loop simulation results. After running closedloop_eval, the statistics of the evaluation results can be obtained by running

python scripts/eval_stats.py -config_dir <path_to_parameter_config_yaml> -model_dir <path_to_model: optional>
python scripts/eval_stats.py -config_dir ./data/configs/params_N14.yaml -model_dir ./data/models/RAIDNET_BC_Jan_21_09_13_20_CA.pt

Furthermore, the RAID-Net model (as a classifier) itself can be evaluated on a test dataset to obtain the confusion matrix and normalized classification loss of the model by running

python scripts/model_eval.py -config_dir <path_to_parameter_config_yaml> -model_dir <path_to_model: optional>
python scripts/model_eval.py -config_dir ./data/configs/model_eval.yaml -model_dir ./data/models/RAIDNET_BC_Jan_21_09_13_20_CA.pt

Combining Datasets (if needed)

If you need to combine two datasets, you can use combine_dataset.py

python scripts/combine_dataset.py -data1_dir <path_to_dataset1> -data2_dir <path_to_dataset2>