Search papers, labs, and topics across Lattice.
This paper introduces UniReflex, a novel framework that enhances pretrained generative imitation learning policies with variable impedance control (VIC) for effective force regulation during contact tasks. By leveraging fast-slow reflex dynamics and intercepting latent representations non-invasively, UniReflex allows for improved contact stability and execution success without the need for retraining the generative models. Experimental results show that UniReflex achieves a significant reduction in latency while maintaining original position accuracy, outperforming traditional joint training methods by 25-66 times.
A non-invasive framework that enhances pretrained generative policies with force control, achieving remarkable improvements in contact stability and execution speed.
Generative imitation learning policies excel at trajectory planning but lack closed-loop force regulation, while directly incorporating force modalities often requires redesigning or retraining the network. We present UniReflex, a universal plug-and-play framework that equips frozen generative policies with variable impedance control (VIC) for contact regulation, guided by force-direction intent collected during demonstration, without further slow-backbone fine-tuning. By non-invasively intercepting deep latent representations from the action head, UniReflex drives a fast reflex network that decouples active force exertion from external interaction response. This scheme predicts normalized anisotropic stiffness directions for directional compliance allocation. Furthermore, UniReflex integrates an adaptive gating mechanism that enables seamless transitions between position-dominant planning and force-dominant execution. Real-world bimanual experiments demonstrate that UniReflex significantly improves contact stability and success rates while preserving original position accuracy. Our approach achieves 25-66x lower per-step backward latency relative to joint training strategies on the evaluated backbones.
Tsinghua AI