SEW Geometric Teleop

Benchmarks

Interactive Retargeting Comparison

Paper Section VI-AFigure 5

Retargeting Accuracy & Performance

Evaluated on LAFAN-1 dataset (upper body motion only) on Unitree G1 robot

Benchmark comparison accuracy chart
SEW-Mimic achieves near-zero error*¹ (near machine precision tolerances) while remaining ~2 orders of magnitudes faster than xr_teleoperate (Unitree) and General Motion Retargeting.
Paper Section VI-CAppendix IFigure 11

Safety Filter

Our safety filter dramatically reduces self-collision of the robot.

Safety filter ablation chart
For the "rolling punch" motion*², the safety filter drastically reduces self-collision rate from 41.3% to 1.4%, while maintaining end-to-end inference rate of 250 Hz. The alignment error sees a small increase of 0.019.
Paper Section VI-BTable IAppendix HFigure 6

Teleoperation User Study

We compare the success of novice teleoperators (n=8) in difficult manipulation tasks.*³

Glass Gap evaluation success rate and completion time
For the Glass Gap task, users show 50.6% higher absolute success rate and 17.2% faster completion time when using SEW-Mimic.
Task completion time comparison
Task completion times between SEW-Mimic and Mink across all tasks. SEW-Mimic shows 15.2% faster mean completion time, which translates to operational efficiency gains during large-scale teleoperation efforts.
Paper Section VI-DAppendix JFigure 12

Policy Learning Impact

SEW-based action representation improves training results of autonomous task policies.

Policy rollout comparison
Policy rollout comparison for models trained on Mink-retargeted vs. SEW-Mimic-retargeted demonstrations.
Action prediction loss comparison
Action prediction error comparison for the trained imitation policies; lower error aligns with improved downstream task performance.
Appendix LFigure 14Figure 15Figure 16

Demonstration Quality Diagnostics

SEW demonstrations show smoother command profiles and less aggressive velocity spikes.

Joint velocity comparison
Per-joint commanded velocity magnitudes for left and right arms; SEW-Mimic trajectories exhibit less aggressive velocity spikes.
Per-joint smoothness comparison
SPARC smoothness comparison per joint showing unanimously smoother action trajectories in demonstrations collected with SEW-Mimic.
Paper Section VI-EAppendix KFigure 7Figure 13

Full-Body Retargeting and Dynamics

SEW-Mimic serves as a drop-in replacement of GMR for the TWIST policy.

SEW-TWIST full-body comparison figure
SEW-TWIST full-body evaluation shows faster inference while maintaining comparable kin2dyn and tracking errors compared to GMR.
Full-body action sequence visualization
A sample motion sequence performed by SEW-TWIST (using SEW-Mimic as full-body retargeter with no fine-tuning of downstream TWIST policy).