Skip to main content

View as Markdown

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

Quadruped series

Explore quadruped modeling, control, reinforcement learning, and multimodal decision-making through projects covering single-joint PD control, full-body gaits, PPO policy training, browser simulation, and language and vision feedback loops.

🗂️ Project overview​

CategoryProjectDescription
Chapter-based courseBuild a quadruped from scratchEight chapters covering MuJoCo actuators, forward and inverse kinematics, Pupper modeling, open-loop gaits, PPO training, LLM control, and visual ball tracking
Browser simulation and terrain experimentsPetoi Quaddle walking simulationControl a four-servo robot in the browser and explore closed-loop linkage adaptation, residual PPO training, and terrain experiments
Language and vision feedbackPetoi Quaddle VLA navigation trainingNavigate using language instructions and camera images: SmolVLA decisions with low-level lookup-table CPG control on a 10GB GPU (L2 navigation 90%, L3 obstacle avoidance 80%, L4 composition 85%)
Cross-simulator validationSim2Sim validationPolicy transfer and consistency checks across physics engines (in development)
Hardware deploymentSim2Real guideCalibration, safety protections, and deployment from simulation policies to quadruped hardware (in development)

🎬 Demos​

CS123 quadruped gait comparison
Build a quadruped from scratch
CS123 simulation · Walk / Pace / Bound / Gallop comparison
PPO gait policy comparison
PPO gait policy training
MuJoCo + Stable-Baselines3 · Closed-loop balance and velocity tracking

Build a quadruped from scratch · Chapter demos​

HFE joint tracking a sine trajectory
1. Actuators and PD control
Joint position feedback and 0.3 Hz sine tracking
Pupper single-leg FK joint sweep
2. Forward kinematics
Three-DoF leg transforms and joint sweeps
Pupper single-leg Raibert stepping
3. Inverse kinematics
Foot trajectory inversion and suspended Raibert stepping
Pupper dropping and holding a standing pose
4. Build a quadruped model
Pupper v3 MJCF, floating base, and stable landing
LLM multi-step language control
7. Language control
LLM function calling for multi-step action sequences
HSV detection and closed-loop ball tracking
8. Visual perception
Onboard camera detection and ball tracking with an RL gait

Petoi Quaddle walking simulation · Scene previews​

Quaddle walking on an infinite checkerboard
Quaddle introduction: browser simulation and residual PPO
Run MuJoCo WebAssembly in the browser to walk, stand, and pause
Quaddle flat-ground stopping and turning
Quaddle experiments: results and failure analysis
Recorded tests and analysis of stopping, turning, slopes, steps, and uneven terrain

Petoi Quaddle VLA navigation · Previews​

L2 language navigation
VLA navigation: train a policy on a 10GB GPU
SmolVLA with temporal ensembling, 90% over 20 episodes of navigating to a red block
Original CPG gait baseline
Low-level control baseline (without VLA)
Open-loop lookup-table CPG gait, upright at 69.8 mm/s

🧭 Course outline​

1. Build a quadruped from scratch (CS123 simulation)​

A Chinese simulation course inspired by Stanford CS123, covering control, kinematics, modeling, gaits, reinforcement learning, language, and vision without requiring physical hardware. The first three chapters also include browser experiments.

ChapterTopicsBrowser experiment
0. Course overviewScope, prerequisites, environment requirements, and a six-week study plan—
1. Actuators and PD controlMotor models, gear ratios, bang-bang control, and joint PD/PID responsesPD experiment
2. Forward kinematicsRigid transforms, homogeneous matrices, DH / URDF conventions, and three-DoF leg FKFK experiment
3. Inverse kinematicsGeometric solutions, Jacobian iteration, and foot trajectory trackingIK experiment
4. Build a quadruped modelPupper v3 URDF/MJCF tree, floating base, and standing parameter sweeps—
5. Gait controlPhase scheduling, swing and stance trajectories, and Trot / Walk / Pace / Bound / Gallop—
6. Policy trainingGymnasium environments, exponential velocity rewards, and PPO gait training—
7. Language controlTool APIs, function calling / ReAct loops, and voice commands—
8. Visual perceptionOnboard cameras, HSV detection, pixel-to-direction mapping, and closed-loop ball tracking—

2. Petoi Quaddle walking simulation​

Based on Petoi's official Quaddle model, with MuJoCo closed-loop linkage constraints, passive torsion springs, and contact parameters for real-time browser interaction.

ChapterTopics
Introduction: make a quadruped walkReal-time MuJoCo in the browser, local mjpython, an infinite ground plane, and a 32-dimensional residual PPO training workflow
Experiments: results and failure analysisRecorded comparisons and reproduction commands for stopping and turning, 3°/5° slopes, 4 mm descents, 2 mm ascents, uneven ground, and candidate running policies

3. Petoi Quaddle VLA navigation training​

Build a VLA pipeline in the Quaddle introduction environment: given "walk to the red block," the robot turns, distinguishes colors, approaches the target, and stops. SmolVLA 450M provides high-level decisions, with lookup-table CPG control and a 10GB GPU memory budget.

ChapterTopics
VLA navigation: train a policy on a 10GB GPUData collection, SmolVLA fine-tuning, and closed-loop evaluation; L2 navigation demos (90%) and capability limits
Experiment notes and v5 roadmapv4/v5/v6 experiments, including negative results; temporal ensembling and 8.2x faster training
Implementation status and plansL1–L4 evaluation progress, task decomposition (L3 obstacle avoidance 80%, L4 composition 85%), and hardware handoff plans