Researchers led by Joohwan Seo introduced ResGAC, a whole‑body controller for humanoid robots. It blends geometric admittance control with a residual reinforcement‑learning layer to cancel unmodelled dynamics. Tested on a real Unitree G1, ResGAC reduced translation and rotation errors compared with the prior SONIC baseline. In a standing peg‑in‑hole trial it succeeded 90% of the time, versus 50% for SONIC.
Why it matters
More accurate tracking lets humanoid robots perform delicate tasks while walking, expanding their use in manufacturing and service settings.
