The International Journal of Robotics Study, Volume 44, Issue 1, Web Page 96-128, January 2025.
Energy viewer (CROWD) can approximate outside joint torque without calling for added sensing units, such as force/torque or joint torque sensing units. Nonetheless, the evaluation efficiency of crowd wears away because of the design unpredictability which incorporates the modeling mistakes and the joint rubbing. Furthermore, the evaluation mistake is considerable when crowd is put on high-dimensional floating-base humanoids, which avoids the approximated outside joint torque from being utilized for pressure control or crash discovery in the genuine humanoid robotic. In this paper, the pure outside joint torque evaluation approach called MOB-Net, is suggested for humanoids. MOB-Net discovers the design unpredictability torque and adjusts the projected signal of crowd, considerably lowering the evaluation mistakes of crowd. The outside joint torque can be approximated in the generalised coordinate consisting of whole-body and online joints of the floating-base robotic with only interior sensing units (an IMU on the hips and encoders in the joints). Moreover, MOB-Net reveals extra durable efficiency for the hidden information contrasted to the end-to-end discovering approach, and the effectiveness of MOB-Net is confirmed via comprehensive simulations, genuine robotic experiments, and ablation research studies. Lastly, different crash managing situations exist to reveal the flexibility of MOB-Net: get in touch with wrench responses control for mobility, crash discovery, and crash response for security.
发布者：Daegyu Lim，转转请注明出处：https://robotalks.cn/mob-net-limb-modularized-uncertainty-torque-learning-of-humanoids-for-sensorless-external-torque-estimation/
