Panasonic's Robot AI Outperforms Rivals in Manipulation Tests

OSAKA, Japan -- Panasonic Holdings and its collaborators have developed T-Rex artificial intelligence model, which manages robotic hands through fingertip sensor and demonstrates a manipulation success rate of 65% as opposed to 35% of the leading competitor’s model.

Panasonic reported Nov. 5 that in 12 manipulation tasks that involved delicate force control T-Rex has performed better than six other AI models. Tasks included flipping pages, moving eggs, squeezing toothpaste, unlocking door with key, and screwing light bulbs in.

T-Rex showed a 65% success rate whereas EgoScale had 35%, π0.5 had 17% and Tactile-VLA had 15% success rates. Other three models had success rates ranging from 3% to 6%.

Unlike other robot AI models that operate with the help of camera and instruction in language format, T-Rex detects pressure, slipping, and contact with objects through tactile sensors.

It comprises two kinds of controls, including a process which uses visual information to perform the action about five times each second. Another control process that uses the information provided by touch performs about 20 times a second and adjusts in case of slipping objects or excessive pressure on the robot.

The average success rate for T-Rex without tactile data was 42%, as reported by Panasonic.

For training, the researchers used 100 hours of robot manipulation data related to about 200 different objects and actions from 22 classes. Also, it has been trained using almost 22,900 hours of human activity videos.

This study involved Panasonic, the University of California, Berkeley, Nvidia, Stanford University, Sapienza University of Rome, and ItalAI. It will be presented at the Conference on Robot Learning in Austin, Texas, Nov. 9-12. Panasonic plans to apply the technology to manufacturing and logistics tasks but has not announced commercial deployments.