Artificial intelligenceBrief
Rhoda AI wants robots to learn from web video, not just teleoperation
A US start-up is building an AI model for several types of robot, trained on public videos rather than teleoperation sessions alone. It is aiming mainly at factories and logistics.
- Also worth watching Rhoda AI: Returns Processing Demo (Rhoda AI, 1 min 51 s, on YouTube)
According to the Korean newspaper Robot Newspaper, which cites The Robot Report, the US start-up Rhoda AI is developing a general-purpose AI model that can be applied to different robots, built from public videos online.
The method breaks with common practice, where an operator controls the robot remotely to collect data. Here, the model first learns how the physical world works from videos.
The model, called DVA (Direct Video Action), analyses past videos along with action and language information. It predicts what comes next as video, then translates that into robot movements.
After this pre-training on a large corpus of videos, the model receives 10 to 20 hours of data specific to the target task.
Its CEO, Jagdeep Singh, says that models trained on a broad set of videos, including natural phenomena and object movements, performed better than those trained only on videos of human manipulation.
Source
- 로봇신문 Robot Newspaper · irobotnews.com
Also worth reading
Illustration, not a photo of the robot
Illustration, not a photo of the robot