Minimalist Compliance Control

Venue: RSS 2026 (Sydney, Jul 13–17) · Session: Manipulation 3 · paper #123 Authors: Haochen Shi, Songbo Hu, Yifan Hou, Weizhuo Wang, Karen Liu, Shuran Song arXiv: 2603.00913 · program page

Summary compiled from the arXiv paper (v1); all numbers quoted from the paper. Trend context: RSS 2026 survey.

Figure 1. (A) The method estimates external wrenches f̂ₑₓₜ directly from motor current or voltage (PWM) signals via a motor torque model and Jacobians, driving a spring–mass–damper (admittance) update of task-space references — no force sensor, no learning. (B) It is embodiment-agnostic across an ARX arm (QDD), a Unitree G1 (QDD), and a LEAP hand (servo). (C) It is plug-and-play with VLM-based, imitation, and model-based policies across tasks like wiping, drawing, scooping, and in-hand manipulation.

Problem

Compliance control is essential for safe physical interaction but is gated by hardware — force/torque sensors or torque-controlled actuators absent from many common platforms (ARX, ALOHA, LEAP, Unitree G1). Recent RL approaches that bypass these sensors suffer sim-to-real gaps, give no safety guarantees (risking large force spikes), and add system complexity.

Method

Minimalist Compliance Control estimates external wrenches from motor current or PWM (voltage) signals — ubiquitous in modern servos and quasi-direct-drive (QDD) motors — using a calibrated motor torque model and the manipulator Jacobian, then feeds the estimate into a task-space admittance controller. It requires no force sensors, no current control, and no learning. A key observation is that compliance depends on correct force sign/direction and reasonable accuracy in the relevant frequency band rather than exact force magnitude, so approximate wrench estimates suffice for stable, responsive behavior. The approach is embodiment-agnostic and plug-and-play with diverse high-level planners.

Results

Validated on a robot arm, a dexterous hand, and two humanoids using VLM-based, imitation, and model-based policies. Estimated forces track an ATI Mini45 reference with mean absolute error 0.69 ± 0.73 N on ToddlerBot (servo, gear ratio > 200:1) and 1.05 ± 1.60 N on the ARX QDD arm. In a quantitative comparison (Table I), the full method achieves position error 15.9 ± 5.1 mm, orientation error 0.048 ± 0.043 rad, and root pitch 0.029 ± 0.012 rad, outperforming baselines UniFP (57.8 mm) and FACET (22.4 mm) and an ablation without wrench estimation (22.5 mm). A contact-rich task reaches an 80% (16/20) success rate, and in-hand manipulation sustains 3.6 cm translation and 27° rotation.

Significance

Democratizes compliant, contact-rich control across embodiments without dedicated force sensing, complementing sensor-light manipulation efforts in Review-Dexterous-Manipulation and Review-Cross-Embodiment.

← Back to RSS 2026 survey · RSS-2026-Papers · Home