113 lines
2.7 KiB
Markdown
113 lines
2.7 KiB
Markdown
# StackForce SimReady Isaac Lab 导出工程
|
||
|
||
这个工程由 StackForce SimReady 导出，可直接用于 Isaac Lab / Isaac Sim 训练。
|
||
|
||
### 直接可复制的训练命令
|
||
|
||
```bash
|
||
conda activate <你自己的IsaacLab环境名称>
|
||
cd <exported_project>
|
||
python -m pip install -e source/stackforce_simready_2foot_lab
|
||
python scripts/list_envs.py
|
||
python scripts/zero_agent.py --task Template-2foot-Direct-v0 --headless --num_envs 8
|
||
python scripts/rsl_rl/train.py --task Template-2foot-Direct-v0 --headless --num_envs 64 --max_iterations 100
|
||
```
|
||
|
||
如果你想打开 Isaac Sim 窗口，把训练命令里的 `--headless` 去掉即可。
|
||
|
||
### 推荐 Isaac Lab / Isaac Sim 环境
|
||
|
||
本导出工程推荐使用下面这套已验证配置：
|
||
|
||
```text
|
||
Python 3.11
|
||
Isaac Sim 5.1.0
|
||
Isaac Lab v2.3.2 / pip 2.3.2.post1
|
||
Torch 2.7.0+cu128
|
||
Torchvision 0.22.0+cu128
|
||
LeggedGym-Ex 0.3.0 提供的 rsl_rl
|
||
```
|
||
|
||
导出包内已包含一键环境脚本：
|
||
|
||
```bash
|
||
chmod +x scripts/setup_stackforce_isaac_lab_sim_env.sh
|
||
./scripts/setup_stackforce_isaac_lab_sim_env.sh
|
||
```
|
||
|
||
脚本默认创建 `env_isaaclab`。如果你想改环境名：
|
||
|
||
```bash
|
||
ENV_NAME=my_isaaclab ./scripts/setup_stackforce_isaac_lab_sim_env.sh
|
||
```
|
||
|
||
|
||
### 训练输出和 checkpoint
|
||
|
||
训练输出在：
|
||
|
||
```text
|
||
logs/rsl_rl/<experiment_name>/<timestamp>/
|
||
```
|
||
|
||
可用下面的命令直接查找：
|
||
|
||
```bash
|
||
find logs -name "*.pt"
|
||
find logs -name "*.pt" | sort | tail -n 1
|
||
find logs -name "policy.onnx"
|
||
```
|
||
|
||
Training saves `model_final.pt` and also attempts to export `exported/policies/policy.onnx`.
|
||
|
||
### 继续训练
|
||
|
||
```bash
|
||
python scripts/rsl_rl/train.py --task Template-2foot-Direct-v0 --resume --load_run <run_dir_name> --checkpoint <model_x.pt>
|
||
```
|
||
|
||
### 播放训练后的策略
|
||
|
||
```bash
|
||
python scripts/rsl_rl/play.py --task Template-2foot-Direct-v0 --num_envs 1 --disable_resets
|
||
```
|
||
|
||
如需加载指定 checkpoint：
|
||
|
||
```bash
|
||
python scripts/rsl_rl/play.py --task Template-2foot-Direct-v0 --checkpoint logs/rsl_rl/<experiment_name>/<timestamp>/model_final.pt --num_envs 1 --disable_resets
|
||
```
|
||
|
||
Regenerate ONNX from a checkpoint:
|
||
|
||
```bash
|
||
python scripts/rsl_rl/play.py --task Template-2foot-Direct-v0 --checkpoint logs/rsl_rl/<experiment_name>/<timestamp>/model_final.pt --num_envs 1 --disable_resets --export_onnx --num_steps 1
|
||
```
|
||
|
||
### 增加自定义 Reward
|
||
|
||
编辑：
|
||
|
||
```text
|
||
source/stackforce_simready_2foot_lab/stackforce_simready_2foot_lab/tasks/direct/2foot/custom_rewards.py
|
||
```
|
||
|
||
在 `compute_custom_reward_terms(env)` 中返回你自己的 reward term。
|
||
然后再编辑：
|
||
|
||
```text
|
||
source/stackforce_simready_2foot_lab/stackforce_simready_2foot_lab/tasks/direct/2foot/2foot_env_cfg.py
|
||
```
|
||
|
||
把：
|
||
|
||
```python
|
||
"custom_reward": 0.0
|
||
```
|
||
|
||
改成非零，比如：
|
||
|
||
```python
|
||
"custom_reward": 1.0
|
||
```
|