Add T800 Huabu AMP

.gitignore

@@ -5,5 +5,4 @@ build/
 *__pycache__/
 *logs/
 *outputs/
-dataset/
-models/
+models/

README.md

@@ -1,153 +1,192 @@
-# EngineAI-Lab
+# T800 Huabu BoxingFrame AMP
 
-**EngineAI Lab** is a python package for training and deploying policies for EngineAI Robots using Isaac Lab and Isaac Sim.
+这个仓库当前主要用于训练 T800 的 boxing-frame 滑步 AMP 策略。
 
-|Training| Sim2Sim |Deploy|
-|--------|--------|--------|
-![training](./docs/training.gif)|![sim2sim](./docs/sim2sim.gif)|<img src="./docs/deploy.gif" height="180"/>
+## 环境与安装
 
-# Structure
+先按 [Isaac Lab 官方文档](https://isaac-sim.github.io/IsaacLab/) 安装兼容版本的 Isaac Sim 和 Isaac Lab，然后在仓库根目录安装本项目：
 
+```bash
+python -m pip install -e .
 ```
-engineai-lab
-├── config
-├── dataset
-│   ├── config
-│   └── data
-├── scripts
-└── source
-    └── engineai_lab
-        ├── algorithms      
-        ├── assets
-        │   └── pm01
-        │       ├── meshes
-        │       └── urdf
-        ├── robots
-        ├── tasks
-        │   └── velocity
-        │       ├── config
-        │       │   └── pm01
-        │       └── mdp
-        └── utils
-```
-
-## QUICKSTART
-
-### 1. Create a Conda Environment
-
-Create and activate a new environment with Python 3.11:
-
-  ```bash
-  conda create -n engineai_lab python=3.11
-  conda activate engineai_lab
-  ```
-
-### 2. Install Prerequisites
-
-- **Install Isaac Sim**
-
-  Follow the official installation guide: [Isaac Lab - Pip Installation](https://isaac-sim.github.io/IsaacLab/main/source/setup/installation/pip_installation.html#installing-dependencies).
 
-    *Since you've already created the engineai_lab environment, follow the guide from "Installing Dependencies" up to (but not including) the "Installing Isaac Lab" section.*
+所有命令都应从仓库根目录运行。训练需要支持 CUDA 的 NVIDIA GPU；显存不足时请降低 `--num_envs`。
+
+## 当前默认设置
+
+- 默认训练轮数：`200000`
+- 默认环境数：`4096`
+- 速度命令坐标系：boxing frame
+- boxing frame yaw offset：`1.03 rad`
+- 命令范围：`lin_vel_x = [-2.0, 3.0]`，`lin_vel_y = 0.0`，`ang_vel_z = [-2.0, 2.0]`
+- 数据集：`dataset/config/t800_dataset_huabu_forward_aligned_with_turn.yaml`，包含 forward-aligned、backward 和 turn 数据
+- terrain：使用 terrain generator，但 level 锁在 `0`，不做 terrain curriculum
+- domain randomization：默认关闭；`AllRand` task 打开默认 T800 随机化
+- policy / critic 网络：`[512, 256, 128]`
+- action distribution：默认 tanh-squashed Gaussian，raw action 范围约为 `[-3, 3]`
+- reset/default pose：tj `rl_amp_huabu/ready.yaml` 的准备姿态
+- entropy coef：tanh `0.008`，Gaussian `0.004`
+- action target：先按关节 scale 加到 default pose，再按 URDF joint limit clip
+- 双臂：live action，但 target 夹在 default `±0.2 rad`，并用强 anchor 拉住
+- AMP：`no_arm_joint_vel`，100Hz 两帧 pair，单帧 36 维
+- final reward：`0.7 * style + 0.3 * task`
+- style weight：`6.0`
+- tracking：`track_lin_vel_xy_boxing_frame_exp` weight `6.0`，sigma `4`；yaw tracking weight `4.0`，sigma `4`
+- `leg_joint_position`：只在低速/静止命令下启用，阈值 `0.15`
+- `foot_position` / `feet_orientation` / `foot_stumble` / `feet_slide`：权重为 `0.0`
+- `feet_contact` / `feet_air_time` / `feet_air_time_dense`：关闭
+- discriminator：每个 PPO iteration 更新一次，mini-batches 为 `4`，lr 跟随 policy adaptive lr，`disc_lr = clamp(0.5 * policy_lr, 1e-5, 1e-3)`
+
+## Task
+
+前进后退优先使用 Task 1 和 2。
+
+| Task | 用途 |
+| --- | --- |
+| `Flat-AMP-T800-Huabu-BoxingFrame-ClampedArmYaw-v0` | Tanh actor，ClampedArmYaw 基线、tj ready reset/default pose、no-arm 100Hz pair AMP、turn 数据、无随机化 |
+| `Flat-AMP-T800-Huabu-BoxingFrame-ClampedArmYaw-AllRand-v0` | ClampedArmYaw + 默认 T800 domain randomization |
+| `Flat-AMP-T800-Huabu-BoxingFrame-ClampedUpperYaw-v0` | ClampedArmYaw 基线，但 torso/双臂/head 都按 default `±0.2 rad` target clip，并用强 anchor 拉住 |
+| `Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-v0` | 默认 yaw 命令采样，只把上半身改成小 action scale + tight target clip |
+| `Flat-AMP-T800-Huabu-BoxingFrame-GaussianSoftUpperYaw-v0` | Gaussian actor + 默认 yaw 命令采样 + SoftUpper |
+| `Flat-AMP-T800-Huabu-BoxingFrame-GaussianSoftUpperYaw-ForwardAligned-v0` | Gaussian actor + 默认 yaw 命令采样 + SoftUpper + forward_aligned 数据集 |
+| `Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-ForwardAligned-CmdMatched-v0` | Tanh SoftUpper + forward_aligned 数据集 + 按 `(vx, wz)` command 相似度采样 AMP expert frame |
+| `Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-ForwardAligned-CmdConditioned-v0` | CmdMatched + discriminator 输入额外拼接 `(vx, wz)`，即 `D(s, s', c)` |
+| `Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-ForwardAligned-HeadingHold-v0` | Tanh SoftUpper + forward_aligned 数据集 + 直线命令 heading hold |
+| `Flat-AMP-T800-Huabu-BoxingFrame-GaussianClampedArmYaw-v0` | 标准 Gaussian actor + ClampedArmYaw |
+| `Flat-AMP-T800-Huabu-BoxingFrame-GaussianClampedArmYaw-AllRand-v0` | GaussianClampedArmYaw + 默认 T800 domain randomization |
+
+## 训练命令
+
+Tanh ClampedArmYaw：
 
-- **Clone & Setup Isaac Lab**
-
-  Clone the repository and switch to the recommended branch:
-
-  ```bash
-  git clone https://github.com/isaac-sim/IsaacLab.git
-  cd IsaacLab
-  git checkout 4df6560e
-  ./isaaclab -i rsl_rl   # Install rsl-rl dependency
-  ```
-
-We highly recommend using the main branch`(4df6560e)` of Isaac Lab, as it can support rsl-rl-lib >= 5.0 and Isaac Sim >= 5.0 .
-
-### 3. Install this Package
+```bash
+CUDA_VISIBLE_DEVICES=0 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-ClampedArmYaw-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name clamped_arm_yaw
+```
 
-Once the prerequisites are set up, install the package in editable mode:
+Tanh ClampedArmYaw + 全随机化：
 
 ```bash
-pip install -e .
+CUDA_VISIBLE_DEVICES=1 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-ClampedArmYaw-AllRand-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name clamped_arm_yaw_allrand
 ```
 
-## Usage
+Tanh ClampedUpperYaw，上半身按双臂的方式约束：
 
-### Supported Robots
-
-This repository currently supports the following environments from the EngineAI Robots family:
+```bash
+CUDA_VISIBLE_DEVICES=1 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-ClampedUpperYaw-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name clamped_upper_yaw
+```
 
-|Robot| Task |Description|
-|--------|--------|--------|
-PM01|`Flat-PM01-v0`|Basic flat-terrain locomotion
-PM01|`Flat-AMP-PM01-v0`|AMP-based motion imitation on flat terrain
+Tanh SoftUpperYaw，默认命令采样：
 
-*More robots and environments are coming soon!*
+```bash
+CUDA_VISIBLE_DEVICES=0 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name soft_upper_yaw
+```
 
-### Training a Policy
+Gaussian SoftUpperYaw，默认命令采样：
 
-```
- python scripts/train.py --task=Flat-PM01-v0  --num_envs 4096 --headless --run_name <name>  
- python scripts/play.py  --task=Flat-PM01-v0 --num_envs 128 --load_run <name> 
+```bash
+CUDA_VISIBLE_DEVICES=2 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-GaussianSoftUpperYaw-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name gaussian_soft_upper_yaw
 ```
 
-### Evaluating a Policy
+Gaussian SoftUpperYaw + forward_aligned 数据集，默认命令采样：
 
+```bash
+CUDA_VISIBLE_DEVICES=2 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-GaussianSoftUpperYaw-ForwardAligned-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name gaussian_soft_upper_yaw_forward_aligned
 ```
- python scripts/train.py --task=Flat-AMP-PM01-v0  --num_envs 4096 --headless --run_name <name> 
- python scripts/play.py --task=Flat-AMP-PM01-v0 --num_envs 128 --load_run <name>
-```
-
-Replace `<name>` with the name of your training run (found in logs/rsl_rl/).
 
-### Deployment
+Tanh SoftUpperYaw + forward_aligned 数据集，AMP expert 按 `(vx, wz)` command 相似度采样：
 
-To deploy a trained policy on real hardware, convert it to the MNN format for efficient inference.
+```bash
+CUDA_VISIBLE_DEVICES=2 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-ForwardAligned-CmdMatched-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name soft_upper_yaw_forward_aligned_cmd_matched
+```
 
-#### 1. Export PyTorch Policy to ONNX
+Tanh SoftUpperYaw + forward_aligned 数据集，command-matched sampling + `D(s, s', c)`：
 
-(Ensure your training script supports ONNX export)
+```bash
+CUDA_VISIBLE_DEVICES=2 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-ForwardAligned-CmdConditioned-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name soft_upper_yaw_forward_aligned_cmd_conditioned
+```
 
-#### 2. Build [MNN-Converter](https://mnn-docs.readthedocs.io/en/latest/start/quickstart_cpp.html?highlight=mnn+converter)
+Tanh SoftUpperYaw + forward_aligned 数据集 + 直线 heading hold：
 
 ```bash
-git clone https://github.com/alibaba/mnn
-cd mnn
-mkdir build && cd build
-cmake .. -DMNN_BUILD_CONVERTER=ON
-make -j8
+CUDA_VISIBLE_DEVICES=2 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-SoftUpperYaw-ForwardAligned-HeadingHold-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name soft_upper_yaw_forward_aligned_heading_hold
 ```
 
-#### 3. Convert ONNX into MNN
+Gaussian actor + ClampedArmYaw：
 
 ```bash
-./MNNConvert -f ONNX \
-  --modelFile path_to_your_policy.onnx \
-  --MNNModel your_policy.mnn \
-  --bizCode MNN
+CUDA_VISIBLE_DEVICES=2 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-GaussianClampedArmYaw-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name gaussian_clamped_arm_yaw
 ```
 
-For detailed instructions on integrating the MNN model with EngineAI robots, see:
-[engineai_robotics_native_sdk](https://github.com/engineai-robotics/engineai_robotics_native_sdk).
+Gaussian actor + ClampedArmYaw + 全随机化：
 
-## Support
+```bash
+CUDA_VISIBLE_DEVICES=3 python scripts/train.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-GaussianClampedArmYaw-AllRand-v0 \
+  --num_envs 4096 \
+  --headless \
+  --run_name gaussian_clamped_arm_yaw_allrand
+```
 
-If you have any questions about using this repository, we're here to help!
+## Play 命令
 
-- **Report Issues**: Found a bug or have a feature request. Please open a new issue on our [GitHub Issues](https://github.com/engineai-robotics/engineai_lab/issues) page.
-- **Email Us**: For general inquiries or collaboration opportunities, feel free to reach out at [info@engineai.com.cn](mailto:info@engineai.com.cn).
+播放 checkpoint 时使用 `--checkpoint`，不是 `--load_checkpoint`。
 
-## License
+```bash
+python scripts/play.py \
+  --task=Flat-AMP-T800-Huabu-BoxingFrame-ClampedArmYaw-v0 \
+  --num_envs 1 \
+  --load_run <run_dir> \
+  --checkpoint <checkpoint_file.pt> \
+  --keyboard \
+  --keyboard_vx_step 0.1 \
+  --keyboard_vy_step 0.0 \
+  --keyboard_wz_step 0.1
+```
 
-EngineAI-Lab is released under [BSD-3 License](LICENSE).
+键盘控制是增量式的：按一次增加或减少一个 step，松手后命令会保持；按 `L` 清零命令。
 
-## Acknowledgement
+## 数据
 
-This repository is built upon the support and contributions of the following open-source projects. Special thanks to:
+公开仓库包含 AMP 配置和部分 `.npz` 动作数据。数据清单位于 `dataset/config/`，路径相对于清单文件解析。
 
-- [**IsaacLab**](https://github.com/isaac-sim/IsaacLab) — The foundational framework for training and running simulation experiments.
-- [**rsl_rl**](https://github.com/leggedrobotics/rsl_rl) — High-performance reinforcement learning library for legged robots.
-- [**AMP_for_hardware**](https://github.com/escontra/AMP_for_hardware) — Implementation of Adversarial Motion Priors (AMP) for sim-to-real transfer.
-- [**BeyondMimic**](https://github.com/HybridRobotics/whole_body_tracking) — Inspiration for project structure and valuable feature implementations.
-- [**MNN**](https://github.com/alibaba/mnn) — Lightweight, high-performance inference engine for on-device deployment.
-- [**engineai_robotics_native_sdk**](https://github.com/engineai-robotics/engineai_robotics_native_sdk) — Official SDK for deploying policies on EngineAI robotic hardware.
+发布或再分发动作数据、机器人 URDF/mesh 之前，请确认你拥有公开这些内容的权利，并在需要时补充独立的数据或资产许可证。

dataset/config/t800_dataset_huabu_forward_aligned_with_turn.yaml

@@ -0,0 +1,4 @@
+motions:
+  - folder: ../data/forward_aligned
+  - folder: ../data/huabu_Fast_B_npz_0609_tran
+  - folder: ../data/npz_turn_0602_1_30hz