Spaces:
Running
Running
Commit History
-
Frontal camera default for policy viewer
Verified signature62230cc -
Fix actuator parameters, foot geometry, and IMU velocity to match training
experientialtech +1: Claude Opus 4.6- Shoulder/elbow joints: Kp 14.25→40.18, effort 25→88 (were using wrist values) - Ankle/waist effort limits: 35→50 Nm (30% more balance torque) - Hip pitch effort: 139→88 to match training - Wrist effort: 13.4→5 to match training - Shoulder/elbow armature: 0.0036→0.0102 (matching Group 2) - Foot geoms: tiny 5mm spheres → proper 10mm capsule arrays (7 per foot) - Added contact exclusion rules from training XML - IMU velocimeter: added omega×r cross-product for offset IMU site - Initial height: 0.793→0.76 matching training init_state Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>ead057c -
Match training control frequency: 50 Hz with dt=0.005
experientialtech +1: Claude Opus 4.6Training uses decimation=4, timestep=0.005, giving 50 Hz policy rate. Previously used 60 Hz with 0.002 timestep. 4 substeps of 0.005s = 0.02s per policy step. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>14f8a2a -
Add interactive WASM policy viewer with ONNX inference
experientialtech +1: Claude Opus 4.6- viewer.html/viewer.js: MuJoCo WASM + Three.js + ONNX Runtime Web viewer for running trained RL policies live in the browser - model/g1_viewer.xml + meshes/: G1 robot model for web simulation - media/: 39 ONNX policy files + 59 NPZ motion references - app.js: "View" button on trained policy cards links to viewer - style.css: viewer button styles - data.json: regenerated with has_onnx flags - Removed stray nested directories Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>e7fed85