Different deployment performance with policy only

[JiahengZhao]

Description

Using loco.onnx and squat.onnx with the same kp/kd configuration on the G1 robot, two issues are observed:

Observed Behavior

Squat policy (squat.onnx)

• Robot's knees appear weak during squat
• Robot tilts to one side (asymmetric behavior)

Locomotion policy (loco.onnx)

• Motor produces loud noise during locomotion
• Behavior suggests kp gains are too large

Configuration

• Policies: loco.onnx, squat.onnx
• Gains: Same kp/kd configuration for both policies
• Hardware: Unitree G1

Expected Behavior

• Squat should be stable and symmetric (no tilting)
• Locomotion should be smooth without excessive motor noise

Questions

• Are the kp/kd values correct for real hardware vs. simulation defaults?
• Should squat and loco policies use different gain profiles?
• Is the knee weakness a gain issue or a policy output issue?