Compliance control — making a robot yield to external forces rather than fighting them — traditionally requires force-torque sensors. These sensors are expensive, fragile, and add calibration burden. Reinforcement learning approaches avoid the sensors but introduce sim-to-real gaps, lack safety guarantees, and add computational complexity. Both paths add something to achieve compliance.
Shi, Hu, Hou, Wang, Liu, and Song remove instead. Their minimalist compliance controller uses only motor current and voltage signals already available from standard servo motors. External wrenches are estimated from the actuator signals and kinematic Jacobians, then fed into a task-space admittance controller. No force sensors. No current control. No learning. The method works across a robot arm, a dexterous hand, and two humanoid robots, pairing with vision-language models, imitation learning, and model-based planning without modification.
The accuracy is sufficient, not optimal. Force estimation from motor current includes friction, backlash, and model errors. But compliance control doesn't need precise force measurement — it needs responsive force estimation, fast enough to yield before damage occurs. The acceptable accuracy threshold is lower than the achievable accuracy from existing hardware signals.
The through-claim is about the location of the bottleneck. The assumption was that compliance requires better sensing. The reality is that the sensing already exists in the actuator — every motor reports its current draw, and current is proportional to torque. The information was always there. What was missing was a controller architecture that could use low-fidelity force estimates productively. The problem was not sensing but utilization. The signal was available; the architecture to exploit it was not.