YAW: Code and Data Appendix

Official implementation and submitted experiment data for “YAW: Predictive Reward Composition via Task-Advantage Alignment for Wheel-Legged Locomotion.” YAW (You Always Walk) learns a bounded, mean-one composition of 22 locomotion rewards and validates the induced virtual PPO update with a fixed task advantage on complementary environment indices.

Authors: Lexing Yang, Houbao Ji, Shaolong Shen, Hongji Huang, Xiangxiao Chen, and Yufeng Ding.

Language: 简体中文

This repository contains:

• the two-wheel, six-action Isaac Lab task;
• YAW and all five comparison methods used in the paper;
• training, checkpoint-selection, independent-evaluation, and plotting entry points;
• CPU regression tests and machine-readable experiment configurations; and
• the submitted training curves and fresh-seed evaluation results.

Development-machine paths, logs, and checkpoints are intentionally excluded. Third-party attributions required for reproducibility and license compliance are retained.

Environment

The submitted experiments used Ubuntu 22.04, Isaac Sim 5.1, Isaac Lab, Python 3.11, PyTorch 2.7, and rsl-rl-lib==3.1.2. Install this project inside a working Isaac Lab environment:

python -m pip install -e source/yaw_bot

The robot USD, URDF, and meshes are included under assets/robots/yaw_bot; no asset-conversion step is needed. The complete machine specification and tested dependency versions are in configs/environment.json. Additional protocol details, including terrain-side friction and robot-side domain randomization, are recorded in PROTOCOL_NOTES.md.

Fast Validation

The CPU-only checks do not launch Isaac Sim:

python -B -m unittest discover -s tests -p "test_*.py"
python scripts/validate_release.py

Reproduce the Paper Experiments

Print the exact commands without launching training:

python scripts/reproduce.py train --dry-run

Train all six methods for three seeds each:

python scripts/reproduce.py train

Select checkpoints by the highest trailing-100 training task reward, evaluate the selected checkpoint with three fresh seeds under the matched automatic command distribution, and recreate the paper tables and figures:

python scripts/reproduce.py select
python scripts/reproduce.py evaluate
python scripts/reproduce.py plot

The plot stage works immediately on the included submitted data and does not require checkpoints or development-machine logs. After a fresh rerun, analyze the new logs and evaluation JSON with:

python scripts/reproduce.py plot --rerun-data

Use --method yaw (or another method name) to run only one method. Device and workload sizes can be overridden for smoke tests; the defaults are the paper values:

python scripts/reproduce.py train --method yaw --num-envs 2 --iterations 2

Paper Methods and Task IDs

Paper method	Task ID
YAW	Template-Yaw-Bot-Predictive-Gated-Direct-v0
Outer-only PPO	Template-Yaw-Bot-Outer-Only-PPO-Direct-v0
Uniform	Template-Yaw-Bot-Uniform-Reward-PPO-Direct-v0
Static	Template-Yaw-Bot-Static-Reward-PPO-Direct-v0
PPO-LIRPG	Template-Yaw-Bot-LIRPG-PPO-Direct-v0
ReLara-PPO	Template-Yaw-Bot-ReLara-PPO-Direct-v0

See PAPER_CODE_MAP.md for the equation-by-equation correspondence and configs/paper_experiments.json for the executable protocol. The 22 atomic reward definitions and scales are available in configs/reward_terms.json.

Result Data

results/ contains the submitted training curves, 18 independent-play evaluation JSON files (six methods by three fresh seeds), per-seed tables, and aggregate statistics used in the paper figures. Checkpoint paths are replaced with release-safe placeholders because checkpoints are not part of this repository.

Citation

@misc{yang2026yaw,
  title  = {YAW: Predictive Reward Composition via Task-Advantage Alignment for Wheel-Legged Locomotion},
  author = {Yang, Lexing and Ji, Houbao and Shen, Shaolong and Huang, Hongji and Chen, Xiangxiao and Ding, Yufeng},
  year   = {2026},
  note   = {Preprint}
}

License and Third-Party Notice

The project code is released under the BSD 3-Clause License. See LICENSE and NOTICE.md. Individual upstream-derived files retain their original copyright and SPDX headers.