Jeonghyeon Kim, Jaesang Kim, Shinsuk Park

Abstract

A stepping controller can take an unnecessary step on a push that a fixed-support hold could have recovered from. Without a force/torque sensor, a humanoid knows a push only through a delayed, noisy estimate, and estimation accuracy or response speed alone don’t say whether a hold decision gets that information in time. This work measures that information requirement directly — by delaying and corrupting the true push-force signal fed to a fixed hold-first controller on a simulated Unitree G1 — then builds an estimator that resolves the pelvis IMU through the floating-base model using the 12 leg torques directly, avoiding the reconstructed-base-velocity step that creates the usual latency/noise trade-off in momentum observers.

Method

• Measured how much delay and noise a push-force estimate can tolerate before a fixed hold-first controller fails to recover without stepping, by directly corrupting the true push-force signal on a simulated Unitree G1
• Resolved the pelvis IMU measurement through the floating-base model using the 12 leg torques, instead of first reconstructing a base velocity — removing the step that creates the latency/noise trade-off in standard floating-base momentum observers
• Used fast, lightly-filtered information to decide whether to hold, and more heavily filtered information to drive the actuation
• Validated on the real Unitree G1 using a load cell as ground truth

Results

• The holdable force fell by about 4% per 10 ms of delay — matching a pendulum model’s predicted rate — and delaying the actuation commands hurt it as much as an equivalent information delay
• Slow force errors up to 30 N and white noise up to 15 N were tolerated; 20 N errors on the timescale of the push itself caused hold failures
• In closed loop, hold-entry latency predicted the normalized holdable force within 0.02 for three of four estimators tested; a tuned momentum-observer baseline reached only 0.85–0.88
• Met its pre-registered recovery criterion more often than a stepping baseline (65 vs. 62 of 80 runs) while taking 46% fewer steps
• On 30–50 N pushes both controllers recovered from, holding in place cut base displacement by 96%, peak force by 44%, and mechanical work by 69%, relative to the stepping baseline
• On the real G1, the estimator detected pushes faster than a momentum-observer baseline, and added delay lowered the holdable force as the simulation study predicted

Skills

Python MuJoCo Model Predictive Control State Estimation