[Lhy6900]

Summary

• add the G1BoxTracking motion-tracking task implementation, large-box scene assets, and box-specific motion loader
• add PPO / APPO owner configs for g1_box_tracking on both mujoco and motrix
• expose the new task through the motion-tracking package exports so it can be selected through task owner YAMLs

Linked Work

• Issue: N/A
• Milestone: N/A

Commands actually run:

UV_CACHE_DIR=/tmp/uv-cache uv run pytest tests/config/test_locomotion_params.py -q -k 'box_tracking and (appo or ppo)'
UV_CACHE_DIR=/tmp/uv-cache uv run python scripts/train_appo.py task=g1_box_tracking/mujoco algo.max_iterations=1 training.no_play=true

[caozx1110]

确保提交的配置中算法及后端组合均能训练，否则不提供对应cfg文件

[Copilot]

Pull request overview

Adds a new object-aware G1 motion-tracking task (“G1BoxTracking”) with supporting motion-loader logic, scene assets, and training owner configs so it can be selected and trained via existing Hydra task-owner YAMLs.

Changes:

• Introduces G1BoxTrackingEnv + configs and exports it through the motion-tracking package.
• Adds BoxMotionLoader / BoxMotionData to load per-frame object (box) state from motion NPZs.
• Adds a new MuJoCo scene with a free-jointed “largebox” plus PPO/APPO task owner YAMLs for Mujoco/Motrix.

Reviewed changes

Copilot reviewed 11 out of 13 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/unilab/envs/motion_tracking/g1/motion_box_loader.py	Adds motion loader that also reads object state arrays from NPZ files.
src/unilab/envs/motion_tracking/g1/box_tracking.py	Implements the G1 box-tracking env, including object terminations, critic observations, and rewards.
src/unilab/envs/motion_tracking/g1/init.py	Exports new env/config and box motion loader types from the G1 package.
src/unilab/envs/motion_tracking/init.py	Exposes new env/config and loader types at the motion_tracking package level.
src/unilab/assets/robots/g1/scene_flat_with_largebox.xml	New scene that includes the robot and adds a free-jointed “largebox”.
src/unilab/assets/robots/g1/g1_sphere_hand.xml	New robot MJCF variant used by the largebox scene.
docs/users/zh_CN/02-simulation-backends.md	Updates backend support matrix to include g1_box_tracking.
conf/ppo/task/g1_box_tracking/mujoco.yaml	Adds PPO owner config for Mujoco backend.
conf/ppo/task/g1_box_tracking/motrix.yaml	Adds PPO owner config for Motrix backend.
conf/appo/task/g1_box_tracking/mujoco.yaml	Adds APPO owner config for Mujoco backend.
conf/appo/task/g1_box_tracking/motrix.yaml	Adds APPO owner config for Motrix backend.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[TATP-233]

确保能训练出来，不然不添加yaml

[Lhy6900]

按这个要求收敛了当前 PR。当前只保留了已验证的 ppo + mujoco owner：conf/ppo/task/g1_box_tracking/mujoco.yaml。未做同等训练验证的 ppo/motrix 和 appo/{mujoco,motrix} owner 已从本 PR 移除，并同步刷新了 support matrix。

本地验证对应命令是：
/home/ubuntu/comunicate/newmain_UniLab/UniLab/.venv/bin/python3 scripts/train_rsl_rl.py task=g1_box_tracking/mujoco

[Lhy6900]

uv run scripts/train_rsl_rl.py task=g1_box_tracking/mujoco training.play_only=true 之后输出play video如下：

play_video.mp4