MicroDuck Movement Policy and Task Development Topic
This topic compiles materials related to reinforcement learning, motion control, physical simulation, task scheduling, and on-device deployment for MicroDuck / OpenDuck Mini V2. The directory is organized in the order of “first running the baseline, then understanding the code, and finally developing tasks”, making it suitable for following the learning process while gradually adding experimental records.
Follow the 8-day learning plan . Study one notebook per day, check for a summary of task demonstrations , or read to see the student recruitment brochure .
There is also a 21-day monthly team learning recruitment notice . Each task is scheduled for 2–3 days. Both sets of plans use the same tutorials and Notebooks, and the original 8-day short-term plan remains intact.
Learning Path
- First, examine basketball balance , to understand the velocity command, 61-dimensional observation, 14-dimensional action, PPO, and contact rewards.
- Then, look at browser physical perturbation , to understand the MuJoCo state, perturbations, recording, and visualization interfaces.
- Move on to stilt walking , and learn about actuator parameters, shape curriculum, action smoothing, and sim-to-real constraints.
- Study swinging rotation and ball balance and FastSAC , to compare different tasks and algorithm interfaces.
- Finally, proceed to ladder climbing . First, run the climbing, table ascent, and standing-up processes, then learn about dual policy switching, rail contact, and curriculum training.
- When running or modifying code is required, enter executable code package .
- When reviewing RDK X5, web demonstrations, and multi-policy scheduling is needed, enter side deployment and web deployment .
- When checking execution according to the live broadcast order is required, open Notebook learning interface .
- On day 8, let the ducklings understand language commands and navigate in different rooms, running LightNav-0 multi-scenario navigation . Use a GPU［graphics processor］ to integrate existing models; no navigation model training is required, and no development board is connected.
Directory Structure
text05-MicroDuck运动策略与任务开发专题/
├── README.md
├── 学习计划.md
├── 招募文案.md
├── 月度组队学习招募文案.md
├── 01-任务资料/
│ ├── 01-MicroDuck篮球平衡强化学习/
│ ├── 02-mjswan-MicroDuck浏览器物理扰动/
│ ├── 03-MicroDuck高跷行走强化学习复现/
│ ├── 04-MicroDuck摆动旋转强化学习复现/
│ ├── 05-MotrixLab-MicroDuck球平衡与FastSAC/
│ ├── 06-MicroDuck梯面攀爬强化学习导读/
│ ├── 07-RDK端侧与网页部署/
│ └── 精选视频/
├── 02-可运行代码/
│ ├── microduck-playground-stilts/
│ │ ├── src/ # 机器人、执行器、任务和奖励
│ │ ├── scripts/ # 训练、评测、导出和推理脚本
│ │ ├── tests/ # 配置和物理不变量测试
│ │ ├── hardware/ # 高跷参数化生成器
│ │ └── docs/ # 训练计划和策略说明
│ └── lightnav-learning/ # 导航服务、场景构建与闭环视频
└── 03-Notebook/ # 8 本任务教程，GPU 导航与 RDK X5/BPU 验收
Notebook Learning Line
The Notebook is recommended to run on an Ubuntu workstation. On Windows, Jupyter can only be accessed via a browser. In this way, MuJoCo, training code, and Python dependencies all exist in the same environment. For motion tasks, the RDK X5 can be connected again, using the development board as an independent BPU inference endpoint. The 8th LightNav-0 navigation tutorial only uses the workstation GPU and does not connect to the development board. The Notebook includes built-in task contracts, model validation, and video output, eliminating the need for an additional layer of “Windows Notebook -> Ubuntu -> RDK” switching.
ONNX is compiled into HBM using the toolchain matching the RDK X5 SDK, and the compilation is performed in Ubuntu Docker. The corresponding task HBM is executed on the RDK side; after the board-side services and simulation workstation are ready, the task video can be regenerated.
Tasks and Learning Content
| Task | Learning Content | Demonstration |
|---|---|---|
| Basketball Balance | Observation and movement, PPO, cyclic network, and contact rewards | Maintaining balance in basketball |
| Browser Physical Disturbance | MuJoCo state, external force interaction, recording, and visualization | Dragging and external force interaction |
| Stilt Walking | BAM executor, shape curriculum, and motion smoothing | Walking on 25 cm / 200 cm stilts |
| Swinging Rotation | Task rewards, ONNX export, and deployment interfaces | Swinging and rotation |
| Ball Balance and FastSAC | FastSAC training, model export, and evaluation | Sphere balance training |
| Ladder Climbing | Dual climbing and standing strategies, rail contact, and curriculum training | GPU inference for ladder climbing, reaching the table, and steady standing |
| RDK X5 / Web Deployment | HBM compilation, BPU inference, TCP, and multi-strategy scheduling | Inference at the board and simulation walking |
| LightNav-0 Visual Language Navigation | Image and language, trajectory prediction, MPC, and multi-scenario experiments | Navigation in residences, living rooms, and corridors |
Supporting Code
02-可运行代码/microduck-playground-stilts/ It is from the local location.
microduck-playground-stilts The reproduction-ready lightweight snapshot generated from the working directory includes the current task code, as well as the added terrain surface, executor, evaluation, and testing materials. No copies are made .venv/Logs, training caches, checkpoints, and generated results; these files are large in size and tied to the local environment, and should not be directly submitted to the main repository.
Each task tutorial explains the model source, training method, and corresponding implementation; the code package retains the upstream license and local experimental instructions.
Video Materials
Each task is displayed separately in the demo overview , including GPU inference, BPU inference, stilt size comparison, full ladder climbing, and multi-task overview. The page uses compressed GIFs; running the Notebook again generates the MP4 video and evaluation report for the corresponding task.
Subsequent Logging Methods
For each experiment, record the task name, code version, checkpoint, number of environments, number of training iterations, random seeds, success judgment, video path, and evaluation results, so as to facilitate reproduction and comparison.