Sapiens reports faster robotics training through partial model updates
The AMW Read
The reported method could change robotics adaptation economics, but its speed and generality remain company claims without independent comparative evidence.
Sapiens reports faster robotics training through partial model updates
Singapore-based Sapiens Technology says its method trains robotics AI in as little as 30 minutes by updating 1 billion to 3 billion parameters instead of an entire model. The company reported training times ranging from 30 minutes to less than a week in experiments using robotic arms and toy robot ducks to pick up objects. It is discussing pilots and scaling with more than 30 robotics and medical companies in the United States, China, and Singapore, according to The Straits Times.
The claim speaks to a central challenge in physical AI: adapting robots to tasks and changing environments. Sapiens says its method lets robots learn after training and respond to variations through text prompts. If those results hold across deployments, faster adaptation could make robotics software more practical for specialized industrial work. The reported experiments do not establish that the same training time or performance applies to every model or task.
For builders and investors, the next evidence to examine is a comparison on the same robot, task, and evaluation criteria. TetraBOT, a potential customer, previously tested conventional training for two months without completing its required tasks and now plans to train more industrial robots using Sapiens' technique. Sapiens also says readable descriptions and execution delays make actions easier to monitor, but the report does not provide independent safety results or commercial deployment outcomes.
