Skip to main content

View as Markdown

Page content converted to Markdown. Use the original page link at the end to explore interactive graphics.

Projects: from simulation to real robot deployment

Choose a task and work through environment setup, training, control, and evaluation with a project, demo, or reproduction study. Follow a sequence of chapters to build a complete system, or choose a standalone experiment to test a method. For questions about principles, look up the related module in foundations. For original papers, data, and code, see the introduction's resources.

Where to start​

  • To build a system from scratch: start with the quadruped project. Its eight chapters cover actuators, kinematics, modeling, gaits, and policy training. Basic Python is required; you can study control and linear algebra alongside the experiments.
  • To learn about arm datasets and hardware operation: read the LeRobot course notes in Chinese, then move on to the SO-101 hardware tutorial. The notes cover Units 0–2 and require basic Python; the hardware tutorial requires a compatible robot.
  • To try a simulation first: choose MicroDuck RL or ACT bimanual manipulation, and set up the required environment.
  • To master wheeled-biped balance and cross-simulator deployment: follow the Flamingo wheeled biped · Isaac Lab course, whose six chapters cover parallel PPO, Constraints as Terminations (CaT), ONNX export, and MuJoCo robustness benchmarking.

Before starting, check the environment and goals. Keep your code, parameters, and results as you work, then use evaluation and failure cases to check your understanding. The project chapters are currently in Chinese; the status labels below describe the available source content.

Directions​

DirectionProjectsSuitable for
Robot arm series9 available, 2 plannedContinuous control, VLA, data collection, imitation learning, and real arms
Quadruped series3 available, 2 plannedReinforcement learning, kinematics, gait control, VLA navigation, and sim-to-real
Biped series5 available, 3 plannedWheeled-biped Isaac Lab + CaT, parallel GPU PPO, and multi-stage curriculum learning
Mobile manipulation3 plannedCombining navigation and manipulation

Available projects​

  • Flamingo wheeled biped · Isaac Lab: build an end-to-end RL locomotion pipeline for an 8-DoF wheeled biped, covering delayed/geared actuators, 4096-env PPO vs CaT, ONNX normalizer fusion, and zero-shot MuJoCo Sim2Sim stress tests (overview and 6 chapters).
  • MicroDuck RL: walking, standing up, and rolling: start with MJCF, BAM actuators, and mjlab task registration; train 18 motion modes in MuJoCo Warp and inspect offscreen evaluation through GIF and MP4 recordings.

AMD projects​

  • AUP Learning Cloud: use Ryzen AI APUs, JupyterHub, Code Server, and ROCm in your browser for exercises, edge inference, and small experiments.
  • MicroDuck RL · AMD ROCm: build ROCm Warp and MuJoCo Warp on a Radeon R9700, fix dynamic broadphase caching, and train biped PPO policies.
  • ACT bimanual training · AMD ROCm: train ACT in BF16 on Radeon GPUs, resume checkpoints, evaluate 20 episodes, and export videos.
  • Explore the Pupper quadruped: the AMD flagship project, with Pupper Locomotion for RL motion policies and Pupper VLA for vision-language-action intelligence.

Simulation projects​

ProjectTechnical pathStatus
Build a quadruped from scratchMuJoCo, PD control, kinematics, gaits, policy training, and perceptionOverview and 8 chapters available
Petoi Quaddle four-servo walking simulationOfficial URDF adaptation, browser controls, periodic gaits, and residual PPO trainingAvailable
MicroDuck RL: walking, standing up, and rollingmjlab, MuJoCo Warp, parallel PPO on CUDA, and 18 motion modesAvailable
Getting started with MuJoCoMJCF, physics simulation, and Python controlAvailable
DDPG InvertedPendulumContinuous-control basics and a DDPG baselineAvailable
DDPG ReacherTarget tracking with a planar robot armAvailable
DDPG PusherContact manipulation and reward designAvailable
ACT bimanual trainingALOHA, imitation learning, ACT, and evaluation over multiple episodesAvailable
π₀.₅ + RECAP: LIBERO reproductionValue models, advantage annotation, ACP, and LIBERO evaluationAvailable
Petoi Quaddle VLA navigation trainingSmolVLA fine-tuning, low-level CPG control, temporal ensembling, and task decompositionAvailable
Sim2Sim validationCross-simulator policy validationIn development
Flamingo wheeled biped · Isaac Lab8-DoF wheeled-biped balance, PPO vs CaT, ONNX export, and MuJoCo Sim2Sim benchmarkingOverview and 6 chapters available

Real robot projects​

The LeRobot notes introduce data and toolchains; the SO-101 project then applies that knowledge to a real robot arm.

ProjectTechnical pathStatus
LeRobot course notes in ChineseRobot learning, datasets, toolchains, and classical roboticsUnits 0–2 compiled
SO-101 + LeRobot hardware tutorialHardware connectivity, safety tests, and action playbackAvailable
ROS2 arm controlROS2 control and arm executionIn development
Sim2Real guideDeploying simulation policies and validating on hardwareIn development