仿真控制台

Taku

Dyna Robotics 的半人形机器人，MuJoCo 3.14 WASM 实时物理仿真。

仿真控制台

Taku

  1. 加载 MuJoCo WASM
  2. 编译物理模型（456 个碰撞凸包）
  3. 加载外观网格（72 万三角形）
  4. 准备就绪
Whole-body control · RL

How the RL controller was built

01 DATA URR points on Taku hands · elbows · chest · base fitted to Dyna's teleop log model rebuilt from dyna.co Training targets · 3.1 h demos · 195 s 3 Dyna episodes, mirrored synthetic · 9,600 s random poses, no self-collision steps · 1,500 s 5–20 cm target jumps Dyna teleop · 72 s operator command + robot response 02 TRAIN · mjlab + PPO reference policy 172 → 21 × 4,096 act obs + reward randomized physics unreachable targets smoothness penalties 3.3 h on 1 GPU 03 EVALUATE 15 cm hand step target ringing ✗ ours ✓ 16 mm · 2.4° hand error, held-out command in simulation (not hardware) policy v1…v5 04 DEPLOY · THIS PAGE drag hands policy · 100 Hz MuJoCo WASM in your browser export targets held out: never trained on
Trained from scratch in simulation, tested on a Dyna command it never saw, then run live in this page. A re-implementation of Dyna-2.1 §3.1.2, not Dyna's controller; all numbers are from simulation.