RoboNaldo
Accurate, stable, and powerful humanoid soccer shooting via motion-guided curriculum reinforcement learning.
CoRL 2026 Oral
🌎English | 🇨🇳中文
RoboNaldo trains Unitree G1 soccer-shooting policies in Isaac Lab. This
repository contains the simulation training stack, staged curriculum presets,
and reference motion used to learn the free-kick policy.
For real-world hardware deployment, export the trained policy to ONNX and run it
with the companion
RoboNaldo Deploy
repository. The demos below show the same free-kick policy in simulation and on
a real Unitree G1.
| Simulation |
Real robot |
free_kick_sim.mp4
|
free_kick_real_small.mp4
|
Free-kick demo: RoboNaldo policy trained in Isaac Lab and deployed on Unitree G1.
Repository Overview
source/whole_body_tracking/: Isaac Lab extension, G1 robot config, tracking task, rewards, observations, commands, and PPO config.
source/whole_body_tracking/whole_body_tracking/tasks/tracking/yaml/: public right-foot task presets.
scripts/rsl_rl/: direct train, play, and evaluation entrypoints.
docs/: detailed setup, task-parameter, and reward references.
News
2026-06 Training and deployment code release.
Getting Started
1. Install Isaac Lab
Install Isaac Sim and Isaac Lab first. This codebase follows the Isaac Lab
extension layout and is intended for the Isaac Lab Python environment.
Recommended baseline:
| Dependency |
Version |
| Isaac Sim |
5.1.0 |
| Isaac Lab |
2.3.2 |
| RSL-RL |
3.1.2 |
| Python |
3.11 |
2. Install the Training Extension
Activate a clean Isaac Lab Python environment, then install this repository's
training extension:
python -m pip install -e source/whole_body_tracking
Isaac Lab 2.3.2 requires the RSL-RL 3.x configuration API. Install the version
used by this release in the Isaac Lab environment before running the scripts:
python -m pip install "rsl-rl-lib==3.1.2"
Run this command after activating the Isaac Lab Python environment. If the
environment is managed by an Isaac Lab checkout, the equivalent command is
<isaaclab-root>/isaaclab.sh -p -m pip install "rsl-rl-lib==3.1.2".
The training, playback, and evaluation scripts target this pinned combination
of Isaac Lab and RSL-RL. Do not apply local changes to Isaac Lab or RSL-RL.
This repository already contains the modified BeyondMimic-style Isaac Lab
extension needed for RoboNaldo, so you do not need to clone or install the
upstream BeyondMimic repository separately. The Python package name is
whole_body_tracking; if another editable extension with the same package name
was previously installed in this environment, reinstall the command above and
verify the active package path with:
python - <<'PY'
import importlib.util
spec = importlib.util.find_spec("whole_body_tracking")
print(spec.origin)
PY
The printed path should point to this repository, for example:
.../RoboNaldo/source/whole_body_tracking/whole_body_tracking/__init__.py.
3. Download Robot Assets
The Unitree G1 description is not committed to this repository. Download it
before running the environments:
mkdir -p source/whole_body_tracking/whole_body_tracking/assets
curl -L -o unitree_description.tar.gz https://storage.googleapis.com/qiayuanl_robot_descriptions/unitree_description.tar.gz
tar -xzf unitree_description.tar.gz -C source/whole_body_tracking/whole_body_tracking/assets/
rm unitree_description.tar.gz
test -f source/whole_body_tracking/whole_body_tracking/assets/unitree_description/urdf/g1/main.urdf
The code resolves this path through whole_body_tracking/assets.py, where
ASSET_DIR points to source/whole_body_tracking/whole_body_tracking/assets.
Do not add an assets/__init__.py; unlike the upstream BeyondMimic setup, this
repository already provides the Python module that owns ASSET_DIR.
The downloaded source/.../assets/ directory is ignored by .gitignore and
should not be committed. The soccer ball is created with Isaac Lab native
SphereCfg, so no separate ball mesh is required.
4. Prepare Motions and Checkpoints
Training requires one retargeted kick motion in RoboNaldo NPZ format. This
repository includes the default right-foot push-kick motion:
The NPZ is ready for training and does not need to be regenerated from CSV. To
replace the default motion, overwrite motions/right_kick.npz with another
RoboNaldo-format NPZ.
Optional: upload the normalized NPZ to a W&B registry:
python scripts/upload_npz.py \
--artifact_path motions/right_kick.npz \
--entity <entity> \
--name right_kick
Training
RoboNaldo uses a staged curriculum. Continue each stage from the previous stage checkpoint.
| Stage |
Purpose |
Right-foot preset |
| Stage 1a |
Plane motion-tracking prior, no task reward |
right_kick/tracking_params.yaml |
| Stage 1b, optional |
Mixed-terrain tracking robustness fine-tune |
right_kick/tracking_mixed_params.yaml |
| Stage 2a |
Small-range static-ball adaptation |
right_kick/task_params_1.yaml |
| Stage 2b |
Wider stationary-ball shooting |
right_kick/task_params_2.yaml |
| Stage 3 |
Dynamic incoming-ball shooting with jump trigger/adaptive sampling |
right_kick/task_params_3.yaml |
For training, use scripts/rsl_rl/train.py directly:
python scripts/rsl_rl/train.py \
--task Tracking-Body-Frame-Flat-G1-v0 \
--motion_file motions/right_kick.npz \
--yaml right_kick/tracking_params.yaml \
--headless \
--logger wandb \
--log_project_name kick \
--run_name right_kick_tracking
For different stages, change the --yaml argument value to switch among different stage presets.
Resume training:
python scripts/rsl_rl/train.py \
--task Tracking-Body-Frame-Flat-G1-v0 \
--motion_file motions/right_kick.npz \
--yaml <yaml_file> \
--resume True \
--load_run <plane_tracking_run_folder> \
--checkpoint model_<iter>.pt \
--headless
It is recommended to use a small policy noise std for Stage 2 and Stage 3 resume runs, so the task policy does not destroy the learned kick prior with excessive exploration.
This release ships right-foot presets and the right-foot reference motion. A left-foot curriculum should use mirrored motion data and change main_foot_name to left_ankle_roll_link.
Use Tracking-Body-Frame-Flat-G1-v0 registry for the paper-style body-frame observation
setup and Tracking-Flat-G1-v0 for external-mocap-style global observation setup.
Multi-mode Stage 3 shooting
The repository also includes the multi-mode stage-3 shooting solution from the
development branch. It combines motion tracking and locomotion commands in one
policy and uses the planner-specific incoming-ball preset:
python scripts/rsl_rl/train.py \
--task MultiMode-Planner-Shooting-Flat-G1-v0 \
--motion_file motions/right_kick.npz \
--yaml multi_mode/stage3_planner_shooting.yaml \
--headless \
--logger wandb \
--log_project_name kick \
--run_name stage3_multimode_shooting
Nominal Stage 3 demo:
play-panorama-574326.mp4
Play and Evaluation
Use scripts/rsl_rl/play.py for playback. Known Stage-2 hot-test run:
python scripts/rsl_rl/play.py \
--task Tracking-Body-Frame-Flat-G1-v0 \
--wandb_path <your_checkpoint_path> \
--yaml right_kick/task_params_2.yaml \
--motion_file motions/right_kick.npz \
--num_envs 1 \
--headless
Use scripts/rsl_rl/eval.py for evaluation:
python scripts/rsl_rl/eval.py \
--task Tracking-Body-Frame-Flat-G1-v0 \
--wandb_path <your_checkpoint_path> \
--yaml <your_yaml_file> \
--motion_file motions/right_kick.npz \
--num_envs 6000 \
--headless
eval.py writes per-episode shot metrics and aggregate accuracy/speed summaries
under logs/rsl_rl/eval/.
Deployment (ONNX Export)
Real-robot deployment expects an ONNX policy from this repo. The exporter writes
policy-obs.onnx with embedded metadata (joint names, PD gains, default poses,
observation/action layout, and motion anchor settings) for
RoboNaldo_Deploy.
| When |
Output |
W&B training (--logger wandb) |
<run_folder>/<run_name>.onnx next to each saved model_*.pt |
play.py playback |
<checkpoint_folder>/exported/policy-obs.onnx |
Run play.py once on the checkpoint you plan to deploy (same --task, --yaml,
and --motion_file as training) to generate the ONNX artifact.
Documentation
Citation
If RoboNaldo helps your research, please consider citing:
@article{robonaldo2026,
title={RoboNaldo: Accurate, stable, and powerful humanoid soccer shooting via motion-guided curriculum reinforcement learning},
author={OpenDriveLab},
journal={arXiv preprint arXiv:2606.11092},
year={2026},
url={https://arxiv.org/abs/2606.11092}
}
Acknowledgement
This repository builds on Isaac Lab (IsaacLab),
BeyondMimic, and
RSL-RL.