Researchers in Norway have developed a three-armed robotic system that uses deep reinforcement learning and tactile sensing to automate the preparation of salmon sashimi.
Researchers at the Norwegian University of Science and Technology have made a robot that can prepare salmon sashimi. This robot is called Sashimi-Bot and has three robotic arms. Each arm performs a specific role. One arm positions and stabilises the salmon. Another arm uses a knife, and the third one uses the chopstick to transfer the slices into the tray. The project also addresses a broader challenge in robotics: handling soft materials.
One of the main challenges in automating sashimi preparation is the way raw salmon changes shape when you touch it. This makes it hard to cut into pieces. To fix this problem the researchers trained the robot in a virtual environment using deep reinforcement learning, an AI technique used to enhance performance through continuous practice. During the learning process the robot practised thousands of movements to learn fish positioning and slice techniques. The researchers then transferred the trained model to the physical robot without further training on real salmon.
To improve cutting accuracy, the system uses a GelSight sensor. The sensor detects pressure changes during cutting using a soft gel surface and an internal camera. The researchers trained the sensor by cutting salmon multiple times. They collected over 12,000 data samples. The final model achieved 95% accuracy and 99% of precision during testing. That feedback allowed the robot to modify its cutting motion before the blade pressed deeply into the cutting board.
The system was evaluated using real salmon loins. The robot cut 34 slices of salmon of thickness varying from 6 to 16 millimetres. During testing, the robot transferred 26 of the 28 slices remaining on the cutting board to a serving tray, with only two transfer failures.
