X-NavDP

Generalizing Navigation Diffusion Policy to Novel Behavior and Embodiments with Group Q-score Reweighted Matching

Tianyu Yang^1,2*  Yiming Zeng^3,2*  Wenzhe Cai^2*  Yuqiang Yang²  Jiaqi Peng^4,2  Hui Cheng³  Jiangmiao Pang²  Tai Wang^2†

¹Fudan University  ²Shanghai AI Laboratory  ³Sun Yat-sen University  ⁴Tsinghua University

Installation | Assets | Training | Evaluation | Results | Citation

Introduction

X-NavDP post-trains a pretrained RGBD camera-based navigation diffusion policy via online reinforcement learning across heterogeneous embodiments. Beyond improving general navigation and obstacle avoidance performance, it gains new capabilities for backing out of traps, long-obstacle detours, and embodiment-aware behavior adaptation. Our contributions to navigation diffusion policy are threefold:

• ⚡ Data-efficient RL post-training. X-NavDP enhances pretrained diffusion policies through efficient large-scale, multi-scene online post-training for stronger general navigation ability.
• ⚒ Structured exploration and stable training. Goal-agnostic diffusion trajectories and Group Q-score Reweighted Matching enable structured exploration, improve training stability, and handle hard cases.
• 🤝 Cross-robot generalization and temporal consistency. Lightweight embodiment modulation and RTC guidance improve cross-robot generalization and temporal consistency, leading to superior post-trained navigation performance.

Installation

Clone NavDP and enter the self-contained X-NavDP baseline:

git clone https://github.com/InternRobotics/NavDP.git
cd NavDP/baselines/x-navdp

Create a Python environment and install the dependencies:

conda create -n navrl python=3.11
conda activate navrl
pip install -r requirements.txt

Install Isaac Sim 5.0.0, Isaac Lab 0.46.2, and isaaclab-rl==0.4.0 following their official instructions. Then install acados following the acados installation guide and Python interface guide. The acados setup is lightweight and mainly requires a CMake build plus the Python interface. Before running training or evaluation, expose the acados libraries and accept the Omniverse EULA:

export ACADOS_SOURCE_DIR=/path/to/acados
export LD_LIBRARY_PATH="${ACADOS_SOURCE_DIR}/lib:${LD_LIBRARY_PATH:-}"
export OMNI_KIT_ACCEPT_EULA=YES
export OMNI_KIT_ALLOW_ROOT=1

Assets

Large assets, datasets, robot USDs, low-level controller checkpoints, and pretrained policy checkpoints are not included in this repository. You need to download these assets manually before running training or evaluation.

The GRScenes100 / N1 evaluation scenes are available from InternRobotics/Scene-N1. This dataset contains sky textures, materials, cluttered easy/hard scenes, InternScenes home scenes, and InternScenes commercial scenes.

The X-NavDP assets are available from InternRobotics/X-NavDP. This repository provides navigation_metadata, robot assets, low-level controller checkpoints, scene_split.json, the NavDP pretrained checkpoint, and the X-NavDP post-trained checkpoint.

After downloading, place or symlink the scene data, metadata, robot assets, and checkpoints in the following layout. The scene root can also be kept elsewhere and passed with SCENE_DIR.

Scene data and navigation metadata x-navdp/ +-- data/scenes/ +-- scene_split.json +-- SkyTexture/ +-- Materials/ +-- cluttered_easy/ +-- cluttered_hard/ +-- internscenes_commercial/ | +-- models/ | +-- Materials/ | +-- scenes_commercial/ +-- internscenes_home/ | +-- models/ | +-- Materials/ | +-- scenes_home/ +-- navigation_metadata/ +-- internscenes_commercial/ +-- internscenes_home/ +-- cluttered_easy/ +-- cluttered_hard/

Code, robot assets, and checkpoints x-navdp/ +-- train.py +-- config/ +-- scripts/ +-- src/ | +-- environment/ | | +-- controllers/checkpoints/ | | +-- humanoid_g1/policy.pt | | +-- quadruped_go2/policy.pt | +-- x_navdp/ | +-- training/ | +-- utils/ +-- eval/ +-- data/robots/ | +-- dingo.usd | +-- unitreeg1.usd +-- pretrain_model/ | +-- navdp_pretrained.ckpt +-- checkpoints/ +-- x-navdp_posttrain.ckpt

Training and evaluation expect SCENE_DIR to point to the scene root. The training config initializes from pretrain_model/navdp_pretrained.ckpt; evaluation examples use the released post-trained checkpoint under checkpoints/. Dingo and G1 load robot USDs from data/robots/, while Unitree Go2 uses the Isaac Lab asset path by default.

Training

The released post-trained model was trained for 24,000 steps on 72 scenes, including 16 clutter scenes and 56 home/commercial scenes.

Single-node 8-GPU distributed training can be launched with:

export SCENE_DIR=/path/to/NavDP/baselines/x-navdp/data/scenes
export NPROC_PER_NODE=8

bash scripts/run_ddp_train.sh \
  --config_file config/x-navdp_config.yaml

For a smaller debug run, reduce the number of processes and environments:

export NPROC_PER_NODE=1

bash scripts/run_ddp_train.sh \
  --num_envs 1 \
  --embodiments dingo \
  --max_steps 100

Training writes periodic evaluation records to the directory passed through --txt_dir. Aggregate these logs into global, per-embodiment, and per-scene success-rate CSV files and plots with:

python scripts/aggregate_success.py ./txt/x-navdp

By default, outputs are written to ./result/<txt_subdir_name>/; for the command above, the output directory is ./result/x-navdp/. Each log row has the form rank,step,episode,success,trainer_success_rate. Home/commercial and clutter results are reported separately. Use --output-dir to override the result directory, or --no-plots to generate CSV files only; run python scripts/aggregate_success.py --help for EMA and milestone options.

Evaluation

Evaluation uses a policy server and an Isaac Lab client.

Start the policy server:

bash eval/scripts/start_policy_server.sh \
  --checkpoint checkpoints/x-navdp_posttrain.ckpt \
  --embodiment quadruped

Run point-goal evaluation:

bash eval/scripts/run_evaluation.sh \
  --config_file eval/config/eval_pointgoal/quadruped_internscene_commercial.yaml

By default, evaluation runs all start-goal samples in the selected scene. Use --num_episodes to evaluate a subset and --max_steps to cap the total simulation steps for smoke tests or debugging.

Evaluation configs are provided under eval/config/eval_pointgoal/, and outputs are written under outputs/evaluation/<embodiment>_<scene_type>/. Each scene directory contains a metric.csv file whose first two columns are success and SPL. To print per-USD scene SR and SPL, run:

python eval/scripts/stat_eval_metrics.py outputs/evaluation/quadruped_commercial

Results

On the project benchmark, X-NavDP improves the overall simulation success rate from 61.20% to 84.28% and improves real-world hard-case success rate from 10% to 65%. See the project results for simulation and real-world comparisons across wheeled Dingo, quadruped Unitree Go2, and humanoid Unitree G1 embodiments.

Citation

@misc{yang2026xnavdp,
  title  = {X-NavDP: Generalizing Navigation Diffusion Policy to Novel Behavior and Embodiments with Group Q-score Reweighted Matching},
  author = {Tianyu Yang and Yiming Zeng and Wenzhe Cai and Yuqiang Yang and Jiaqi Peng and Hui Cheng and Jiangmiao Pang and Tai Wang},
  year   = {2026}
}

License

Original X-NavDP code is released under the MIT License. Vendored and external dependencies retain their upstream terms; see THIRD_PARTY_NOTICES.md.

Acknowledgement

This project builds on Isaac Sim, Isaac Lab, acados, and NavDP. X-NavDP is distributed as a self-contained baseline in the InternRobotics/NavDP repository.