behavioral-score-diffusion

Behavioral Score Diffusion (BSD)

Official implementation for “Behavioral Score Diffusion: Model-Free Trajectory Planning via Kernel-Based Score Estimation from Data” (CDC 2026).

[Project Page]

Overview

BSD is a training-free and model-free diffusion planner that computes the diffusion score function directly from a library of trajectory data via kernel-weighted estimation. At each denoising step, BSD retrieves relevant trajectories using a triple-kernel weighting scheme (diffusion proximity, state context, goal relevance) and computes a Nadaraya-Watson estimate of the denoised trajectory.

Key results:

• 98.5% of model-based baseline reward without any dynamics model
• 18–63% improvement over nearest-neighbor retrieval
• Safety shielding transfers directly from model-based methods

Installation

1. Install Docker and NVIDIA Container Toolkit (CUDA):

docker run --rm --gpus all nvidia/cuda:12.0.0-base-ubuntu20.04 nvidia-smi

1. Build and start the container:

docker compose build
xhost +local:docker
docker compose up -d

1. Enter the container:

docker exec -it jax_dev bash

Quick Start

Run BSD planner

from mbd.planners.bsd_planner import BSDPlannerConfig, run_bsd_diffusion
from mbd.planners.mbd_planner import MBDConfig
import mbd

# Configure
config = MBDConfig(env_name="tt2d", case="parking", num_trailers=1)
env = mbd.envs.get_env(config.env_name, case=config.case, dt=config.dt,
                        H=config.Hsample, num_trailers=config.num_trailers)

# Run BSD
run_bsd_diffusion(args=config, env=env)

Reproduce experiments

# Run all BSD experiments (4 systems x 4 conditions x 50 trials)
python experiments/run_bsd_experiments.py

# Run specific system
python experiments/run_bsd_experiments.py --env_name tt2d

Collect trajectory data

python -m mbd.data.collect_data --env_name tt2d --n_trajectories 1000

Supported Systems

System	State Dim	Description
kinematic_bicycle2d	3D	Kinematic bicycle (x, y, theta)
tt2d	4D	Tractor-trailer (x, y, theta1, theta2)
n_trailer2d	5D	N-trailer with additional joint
acc_tt2d	6D	Acceleration-controlled tractor-trailer

Project Structure

mbd/
├── planners/
│   ├── mbd_planner.py       # Model-based diffusion (baseline)
│   └── bsd_planner.py       # Behavioral score diffusion (ours)
├── scorers/
│   ├── behavioral_score.py   # Kernel-weighted score estimation
│   └── bandwidth_schedule.py # Adaptive/fixed bandwidth
├── data/
│   ├── trajectory_dataset.py # Dataset loading and management
│   └── collect_data.py       # Data collection from MBD
├── robots/                   # Vehicle dynamics models
├── envs/                     # Environment setup & visualization
└── utils.py                  # JAX utilities, rollout, animation

Citation

If you find this repository useful, please consider citing our paper:

@inproceedings{bsd2026,
  title     = {Behavioral Score Diffusion: Model-Free Trajectory Planning
               via Kernel-Based Score Estimation from Data},
  author    = {Li, Shihao and Li, Jiachen and Xu, Jiamin and Chen, Dongmei},
  booktitle = {IEEE Conference on Decision and Control (CDC)},
  year      = {2026},
}

Acknowledgments

This implementation builds upon the Safe-MPD codebase by Kim et al. and the Model-Based Diffusion framework. We thank the authors for making their code publicly available.