Singapore's Sapiens Says It Can Train Robot AI in 30 Minutes
A Singapore startup says it has cut the time needed to teach a robot a new task from months to as little as 30 minutes, by training part of a model rather than all of it.
Sapiens Technology retrains between one billion and three billion parameters instead of whole models that run to hundreds of billions, so a base model can be adapted to a new task by adjusting only the part that carries it. The company says this takes 30 minutes to under a week, against weeks or months of standard robot training.
And the commercial interest is already there. Sapiens is in talks with more than 30 robotics and medical firms in the United States, China and Singapore, including TetraBOT Electronic Technology and Shanghai Xinwei Medical Technology.
TetraBOT is the clearest case. Its robots turn knobs, flip switches and press buttons on substation equipment, and two months of conventional training had failed to make them perform. After testing the Sapiens method, the firm plans to train more of its fleet.
Sapiens demonstrated the speed on commercial robot arms and off the shelf toy ducks picking up objects, and says a trained robot can adapt when the task shifts and can take text prompts describing variations rather than new code.
But the claim that matters most in a factory is auditability, and here Sapiens has an answer worth testing. Bruce Yang, the chief executive, says the AI's reasoning is distilled into readable language and a deliberate delay sits between planning an action and executing it, so a human or a monitoring system can stop it.
Shanghai Xinwei, a maker of medical implants, is using the technique differently, to forecast inventory gaps and model how consumables and distributors connect, a task where off the shelf models struggled with the relationships in its procurement pipeline.
The method is a training strategy applied to existing models, including the company's own Agnes AI, rather than a new architecture, which means the results depend on how well partial retraining holds up outside demonstrations with toy ducks and laboratory arms.
So the test is a production line rather than a demo, and the evidence to watch is a customer publishing results from a robot working a live substation instead of a bench.
Sapiens Technology says it trains robot AI in 30 minutes to under a week
The method retrains one to three billion parameters rather than full models
It is in talks with more than 30 robotics and medical firms in the US, China and Singapore
TetraBOT plans to train more substation robots after conventional training failed
Sapiens says its reasoning is readable and actions can be paused before execution
Source: The Straits Times
