Answers are grounded in this paper and cite their sources.
Robotic meat cutting must adapt to large variations in tissue stiffness, which position-controlled motion cannot do directly, whereas admittance control typically requires a force/torque (F/T) sensor that complicates deployment in food processing because of washdown, calibration drift, and cost. In this paper, we present a vision-based control switching framework for robotic meat cutting that combines imitation learning with sensorless force estimation on a 6-DoF manipulator. An Action Chunking with Transformers (ACT) policy driven by a hand–eye camera performs the free-space approach, and a SAM2-based visual proximity trigger on a side-view camera detects when the tracked knife and meat come into close proximity in the image. On this trigger, the system switches to admittance control driven by a cutting force estimated proprioceptively from the joint-4 motor current and joint velocities; the estimator is trained against F/T ground truth, but needs no F/T sensor at inference. In leave-one-episode-out evaluation over 25 pork belly episodes, the estimator predicted the uncompensated F/T magnitude with R2=0.96 and 4.4 N RMSE; relative to the contact force, the deployed estimate showed a systematic underestimate of about 14 N in the training session, which persisted in later sessions with a larger error scatter. On a real robot, the proposed method succeeded in 12 of 15 pork belly cutting trials (80.0%), whereas the ACT-only baseline succeeded in just 4 of 15 (26.7%; Fisher exact test, p=0.009).
No takes yet. Share an insight, caveat, or question.
Yoo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: