WORLDTECHWrist-target construction decides whether humanoid arm joints reach their limits, arXiv study finds
A preprint posted to arXiv contends that the way a humanoid robot's desired wrist position is constructed from human motion, rather than posture matching alone, largely determines whether its arm reaches permitted joint-angle limits. The authors test the claim on a Unitree H1-2.
How the wrist target is assembled
Motion retargeting carries a recorded human movement over to a robot and computes the joint motion that reproduces it. The study focuses on the desired wrist position, which is built from demonstrated arm directions together with the robot's segment lengths. Consequently, the shoulder-to-wrist distance shifts with the demonstrated angle between the segments, even when no posture-matching preference is applied.
The reach-normalized alternative keeps the demonstrated direction and rescales the distance instead. Both constructions use the same processed demonstrations and the same joint-angle method, so the effect of target construction can be isolated.
Results in simulation and on hardware
In a forward-reach sweep, the reconstructed target placed a joint angle at a bound in 10 of 20 repetitions, while the reach-normalized target did so in none. Across 25 recordings from one participant, the counts were 8 and 6 on the reaching arm and 14 and 5 on the supporting arm. A second participant repeated the bricklaying and lifting comparison.
During hardware replay, a joint reached its commanded bound in 12 of 16 direction-matching arm records and in none of the 16 normalized records. Within the eligible records, tracking error and executed torque at those frames were both higher than elsewhere in the same run.
Limits of radial adjustment
The authors identify cross-body reach and arm-to-arm clearance as the conditions that bound how far the target can be adjusted radially, marking where normalization runs out of room.
Implications for retargeting
The findings support treating target construction and posture matching as two distinct retargeting decisions. The authors argue both should be evaluated alongside the wrist positions and clearance that manipulation requires.
Open questions
- Will the reported counts hold across more participants and other robot platforms?
- How well does the reach-normalized target preserve the intended task when cross-body reach is required?