Answers are grounded in this paper and cite their sources.
Contact-rich manipulation demands stable, low-impact interaction with uncertain environments. Reactive diffusion policies (RDP) achieve strong task success, yet their high-frequency corrective actions can inject energy into the contact, producing large contact forces and instability under disturbance. We propose PaRDP (Passivity-Aware Reactive Diffusion Policy), which augments a slow-fast visual-tactile diffusion base with three lightweight, provably passivity-bounded modules: (i) a Passivity-aware Interaction Frame & Hybrid Force-Position (PIF-HFP) module that selects free-motion degrees of freedom in an estimated interaction frame; (ii) a Passivity-Bounded Fast Force Correction (PB-FFC) shell that projects out energy-increasing corrections through an energy tank with a bounded injection allowance; and (iii) Stability-Conditioned Training (SCT), a training-time regularizer that penalizes net energy injection. We prove that the closed-loop system remains stable whenever the environment is passive. On RDP/TactAR, our primary benchmark, PaRDP improves the All score by $$+0.02$$ to $$+0.05$$ over RDP(Force) and reduces peak contact force by up to $$34%$$ while adding under $$7%$$ inference overhead; we additionally list published ForceVLA-Data and RH20T baselines as external reference context for force-aware and cross-dataset settings. Ablation and sensitivity studies confirm the contribution of each module.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: