Trained on 130 teleoperated demonstrations, an ACT policy runs at 15 Hz on a Jetson AGX Orin to detach the milling cutters without damaging them. The data, weights and code are published as open source.
To keep a strawberry saleable, it's best not to touch it: the ideal picking is to pinch and detach the peduncle, a stem 1.3 to 3 mm in diameter, the same colour as the foliage and often hidden by it. Conventional approaches struggle to link vision, movement planning and precise contact in such an environment.
One team proposes a copycat learning chain that addresses these three problems in one piece. Human operators first remotely operated a low-cost robotic arm to provide 130 demonstrations; an Action Chunking with Transformers (ACT) policy, with a ResNet-50 for vision, was then trained over 100,000 iterations, with the loss dropping from 2.84 to 0.091. The policy runs autonomously at 15 Hz on a Jetson AGX Orin onboard board.
The interest is also in openness: the dataset, the model weights and all the code and deployment instructions are published on Hugging Face and GitHub, so that the agricultural robotics community can reproduce and extend the work.
Read the full article on the original source
Go to the source →