huggsook/connect-ai-microduck-rl

🦆 connect-ai : Microduck Bipedal Robot Reinforcement Learning Policies

🎬 Preview Demonstration (Direct MuJoCo Simulation Showcase)

📌 Preview Video Actions Timeline (영상 속 5대 시뮬레이션 동작 순서)

1. [0:00 ~ 0:03] 🚶 Omnidirectional Walking: 전진 걷기, 가속 및 방향 전환 보행 (Velocity Tracking @ 50 Hz)
2. [0:03 ~ 0:06] ⚽ Dynamic Ball Kick: 왼발을 들어 전방의 공을 힘차게 슈팅 (Precision Ball Strike)
3. [0:06 ~ 0:09] 🌀 Roulade (Forward Roll): 앞으로 360도 회전 구르기 후 두 발로 완벽 착지
4. [0:09 ~ 0:12] 📦 Ground Pick: 부리로 바닥을 정확히 터치한 후 직립 복귀 (Phase-Coordinated Crouch)
5. [0:12 ~ 0:16] 🪑 Smooth Sit ↔ Stand: 부드럽게 앉았다가 다시 서는 자세 전환 제어

---

📦 Included ONNX Policies (policies/)

All policies adhere to a unified 61-dimensional observation contract (obs[1, 61] -> actions[1, 14]):

Policy File	Task / Role	Description
alpha_walking.onnx	Velocity & Walking	Robust omnidirectional walking, acceleration, backward & strafing
alpha_stand.onnx	Stand & Body Control	Stable posture holding with 6-DOF trunk & head pose command tracking
alpha_sitstand.onnx	Sit ↔ Stand	Smooth sit down and stand up transitions
ball_kick_left.onnx	Ball Kick (Left Foot)	Dynamic ball strike with left leg
ball_kick_right.onnx	Ball Kick (Right Foot)	Dynamic ball strike with right leg
roulade.onnx	Roulade / Forward Roll	Forward roll over head with momentum-assisted standing recovery
alpha_ground_pick.onnx	Ground Touch / Pick	Crouch down to touch the ground with the beak and return to stand
roller.onnx	Roller Skating	Locomotion on passive roller skates
roller_crouch.onnx	Roller Crouch	Low-center crouch gliding on roller skates

---

🕹️ Interactive Controls (Inference Simulation)

• Arrow Keys (↑ / ↓ / ← / →): Walk forward, backward, left/right strafe
• A / E: Continuous left/right turning
• X: 🔄 180° Turn Around (U-Turn)
• SPACE: Halt / Stand Still
• P: 🔥 Apply Random Push Impulses (Fall recovery & impact testing)
• U / O: 🔄 Instant Stand-Up Recovery
• K / L: ⚽ Kick with Left / Right foot
• C: ⚽ Reset ball to scene center (0, 0)
• R: 🌀 Forward roll (Roulade)
• Y: 🪑 Sit down / Stand up toggle

---

🚀 Quick Start & Deployment

1. Run inference directly with MuJoCo:

python scripts/infer_policy.py \
  --walking policies/alpha_walking.onnx \
  --standing policies/alpha_stand.onnx \
  --sitstand policies/alpha_sitstand.onnx \
  --kick-left policies/ball_kick_left.onnx \
  --kick-right policies/ball_kick_right.onnx \
  --roulade policies/roulade.onnx \
  --new-cmd-obs

2. Observation Contract (61D Vector)

• [0:3]: Base angular velocity (gyroscope, rad/s)
• [3:6]: Projected gravity vector R^T * [0, 0, -1]
• [6:20]: Joint positions relative to default standing pose (14 DOF)
• [20:34]: Joint velocities (14 DOF)
• [34:48]: Previous action history (14 DOF)
• [48:61]: 13D Command Vector: [vx, vy, vyaw (3), head_pose (4), body_pose (6)]

---

📄 License & Credits

• Project: connect-ai Microduck RL
• Base Architecture: Pollen Robotics Microduck & microduck_rl
• License: Apache-2.0