g1-teleop/server/calibration_log.md

Joint Calibration Log

Using simple basis change: R_delta_robot = R_basis @ R_delta_xr @ R_basis.T Decomposition order for shoulders: YXZ Euler (matches URDF: pitch(Y), roll(X), yaw(Z))

l_shoulder_pitch (joint 15) - DONE

• Robot motion: sine ±0.6 rad, period 8s (arm forward/backward)
• XR parent-child: chest -> l_upper
• Result: Y component (YXZ Euler) correlates positively with robot motion
  • Robot +0.57 → XR Y ≈ +0.3 to +0.6
  • Robot -0.58 → XR Y ≈ -1.0 to -1.1
• Sign: SIGN_SH_L[0] = +1 (CORRECT, same sign)
• Note: X (roll) had constant negative offset (~-0.4 to -0.8), likely natural arm rotation during pitch motion

l_shoulder_roll (joint 16) - DONE

• Robot motion: sine ±0.6 rad, period 8s (arm out to left side)
• XR parent-child: chest -> l_upper
• Result: Z component (YXZ Euler) correlates with robot roll, NOT X
  • Robot +0.60 → Z ≈ +0.93 to +1.23
  • Robot -0.60 → Z ≈ +0.05 (can't push arm into body)
• Sign: +1 (same sign)
• Important: This means YXZ Euler order may be wrong, or axis assignment needs swapping (roll=Z, yaw=X?)

l_shoulder_yaw (joint 17) - DONE

• Robot motion: sine ±0.4 rad, period 8s (upper arm twist inward/outward)
• XR parent-child: chest -> l_upper
• Result: X component (YXZ Euler) correlates negatively with robot yaw
  • Robot +0.40 → X ≈ -0.90 to -0.98
  • Robot -0.40 → X ≈ -0.001 to -0.09
• Sign: -1 (inverted)

LEFT SHOULDER SUMMARY

• Pitch → Y component of YXZ, sign +1
• Roll → Z component of YXZ, sign +1
• Yaw → X component of YXZ, sign -1
• Effective Euler order is YZX, not YXZ!
• Code fix: as_euler('YZX') gives (pitch, roll, yaw) directly, with yaw negated

l_elbow (joint 18) - DONE

• Robot motion: sine ±0.7 rad, period 8s (elbow bend/extend)
• XR parent-child: l_upper -> l_lower
• Result: Z component (YXZ Euler) correlates positively with robot elbow
  • Robot +0.69 → Z ≈ +0.82 to +0.92
  • Robot -0.69 → Z ≈ -0.65 to -0.93
• Sign: +1 (same sign)
• Note: Currently code extracts Y[0] for elbow, should use Z[2]

r_shoulder_pitch (joint 22) - DONE

• Robot motion: sine ±0.6 rad, period 8s (right arm forward/backward)
• XR parent-child: chest -> r_upper
• Result: Y component (YXZ Euler) has weak positive correlation
  • Large constant offset (~0.8 rad) - user's arm drifted forward
  • Robot +0.6 → Y ≈ 0.92, Robot -0.6 → Y ≈ 0.76 (difference ~0.16)
• Sign: +1 (same as left, weak signal)
• Note: Right arm Quest tracking is significantly weaker than left

r_shoulder_roll (joint 23) - DONE

• Robot motion: sine ±0.6 rad, period 8s (right arm out to side)
• XR parent-child: chest -> r_upper
• Result: Z component (YXZ Euler) correlates (weak/intermittent signal)
  • Robot -0.60 (arm out) → Z ≈ -0.82 to -0.96 (strong cycles)
  • Robot +0.60 (arm in) → Z ≈ +0.07
• Sign: +1 (same sign)
• Note: Very noisy, right arm Quest tracking weak

r_shoulder_yaw (joint 24) - SKIPPED

• Assumed same as left: X component, sign -1 (right arm tracking too weak for subtle yaw)

r_elbow (joint 25) - DONE

• Robot motion: sine ±0.7 rad, period 8s (right elbow bend)
• XR parent-child: r_upper -> r_lower
• Result: ALL axes near zero (max ±0.07 despite ±0.7 robot motion)
• Quest right arm tracking is extremely weak - barely detects elbow bend
• Assumed: Same mapping as left elbow (Z component, sign +1)

EXTENT TUNING

l_shoulder_pitch (joint 15)

• Backward (positive): +0.8 rad (+45.8 deg) — OK
• Forward (negative): -2.4 rad (-137.5 deg) — OK
• Note: positive = backward, negative = forward for this joint
• Settings: amplitude=1.6, offset=-0.8

l_shoulder_roll (joint 16)

• Inward (negative): ~0 rad (0 deg) — just touching body, no clipping
• Outward (positive): +1.585 rad (+91 deg) — arm straight out to side
• Settings: amplitude=0.8, offset=0.785
• Note: needs outward bias to avoid body clipping

l_shoulder_yaw (joint 17)

• Inward (positive): +1.15 rad (+66 deg)
• Outward (negative): -1.15 rad (-66 deg)
• Settings: amplitude=1.15, offset=0.0
• Note: clips into body at rest pose but OK since other joints move arm away in practice

l_elbow (joint 18)

• Extend (negative): -0.8 rad (-45.8 deg)
• Bend (positive): +1.33 rad (+76.2 deg)
• Settings: amplitude=1.065, offset=0.265
• Note: positive = bend, negative = extend (opposite of label in joint_test.py)

l_wrist_roll (joint 19)

• Palm down (positive): +1.35 rad (+77.3 deg)
• Palm up (negative): -1.35 rad (-77.3 deg)
• Settings: amplitude=1.35, offset=0.0
• Note: positive = palm down, negative = palm up (opposite of label)

l_wrist_pitch (joint 20)

• Wrist down (positive): +0.8 rad (+45.8 deg)
• Wrist up (negative): -0.8 rad (-45.8 deg)
• Settings: amplitude=0.8, offset=0.0
• Directions confirmed correct

l_wrist_yaw (joint 21)

• Inward (negative): -1.28 rad (-73 deg)
• Outward (positive): +0.5 rad (+29 deg)
• Settings: amplitude=0.89, offset=-0.39
• Note: negative = inward, positive = outward

OVERALL AXIS MAPPING SUMMARY

After R_basis @ R_delta @ R_basis.T + as_euler('YXZ') → (Y, X, Z):

Shoulder (chest → upper_arm) — parent roughly upright:

• Robot pitch (Y joint) = Y component [0], sign +1
• Robot roll (X joint) = Z component [2], sign +1
• Robot yaw (Z joint) = X component [1], sign -1

Elbow (upper_arm → lower_arm) — parent hangs down:

• Robot pitch (Y joint) = Z component [2], sign +1
• Note: axis mapping differs from shoulder because parent orientation differs

Why axes differ per joint:

The basis change R_basis @ R_delta @ R_basis.T treats the delta as if in XR world-aligned frame, but the delta is in the PARENT's local frame. When the parent hangs down (upper arm), its local axes are rotated ~90° from world, causing the axis mapping to shift.