The 75 percent figure is 45 trials
Researchers at the Istituto Italiano di Tecnologia, the University of Genoa and TU Delft have posted HIL-HARC, a real-world reinforcement-learning system that separates the robot arm's continuous motion from its gripper commands while training both against a shared, two-part critic. The paper evaluates tennis-ball placement, banana placement and pot reset on physical arms. After 160 minutes of training, it reports 80 percent, 90 percent and 55 percent success respectively, against 60 percent, 60 percent and 0 percent for HIL-SERL.[1]
The useful comparison is smaller than the headline, but still material. Each physical task has 20 evaluation episodes. The reported rates therefore convert to 45 successes from 60 attempts for HIL-HARC, compared with 24 for the baseline: 21 additional completed attempts, not proof of a general manipulation system. The fourth table result, 95 percent for a Unitree G1 block-relocation task, is simulation and should not be added to the physical tally.[1,2]
The hard operating work is still outside the policy
The paper's real-world training loop is not a hands-off learning process. It starts each task with 20 keyboard or joystick demonstrations, uses corrective human interventions when behavior degrades, resets the robot with scripted motions and says that environment resetting is manual. The authors estimate 69, 76 and 111 expert-equivalent episodes for their three physical tasks, versus 80, 102 and 132 for HIL-SERL. That is 256 rather than 314 equivalent episodes, a reported reduction of 58, but it does not remove the operator or reset workload.[1,3]
This changes the practical question. HIL-HARC is evidence that a shared critic and reward decomposition can improve a narrow, physically trained tabletop loop under wider task randomization than its chosen baseline. It is not evidence that a factory can leave a robot to learn unattended: the study has three tabletop tasks, 20 final trials per task, author-reported outcomes and no independent reproduction. The current project page supplies evaluation and training videos, but its resource section offers no code or dataset link for an outside group to repeat the comparison.[1,2,3]
The next meaningful checkpoint is a versioned release with the policy, reward logic and reset protocol, followed by longer independent runs that count every intervention, reset and failure across new objects and workcells. Until then, the result narrows one important problem in robot learning: it improves what happens inside a supervised test loop, not the operations required to keep that loop running.[1,2]