Singapore start-up Sapiens cracks AI training bottleneck, makes waves overseas

SINGAPORE – Singapore artificial intelligence start-up Sapiens Technology has cracked one of robotics’ biggest bottlenecks by slashing machine training time from months to as little as 30 minutes.

Its novel technique – which can train all AI models using only a fraction of their parameters rather than the whole system – is gaining attention overseas.

Sapiens is in talks with more than 30 robotics and medical firms across the United States, China and Singapore to pilot and scale its breakthrough. These potential customers include China’s TetraBOT Electronic Technology and Shanghai Xinwei Medical Technology.

“Our technique will also shorten the time it takes for AI to be adapted for different industries and situations,” Bruce Yang, chief executive officer of Sapiens Technology, told The Straits Times.

Specifically, the technique trains AI on common elements applicable to any scenario, such as understanding its operating environment and autonomously improving its performance. But for different scenarios, what changes in the training are the testing environment, the exact tasks, and how the AI is evaluated.

The method works by training only one billion to three billion “parameters” – the adjustable connections within an AI system – rather than an entire model, which can span hundreds of billions or even trillions of parameters.

This technique allows the machines to be trained in 30 minutes to less than a week, much shorter than the weeks to months typically required to train a robotics application.

Sapiens proved these speeds in experiments, such as using commercial robotic arms and off-the-shelf robot ducks to pick up objects.

After the robot’s AI has been trained, the machine can adapt and learn on the fly even if there are changes in the tasks it has been trained on. The robot’s human users can also use text prompts to instruct the robot to carry out variations on the trained tasks, without having to send or code very detailed and specific instructions.

To ensure that the robot does not take unintended actions after training, Yang said Sapiens ensures that the training and thinking process of the robot can be audited and monitored – unlike other AI training methods.

The AI’s logic is distilled into natural, readable language so humans can easily audit its thinking. A custom delay between the time the AI plans an action and when it executes the action also allows humans or AI monitoring systems to halt any unintended or rogue behaviour.

The new technique works with all AI models including Sapiens’ own Agnes AI developed in Singapore.

One potential customer, TetraBOT, which earlier tested Sapiens’ technique, is now planning to train more of its industrial robots.

TetraBOT’s robots turn knobs, flip switches and press buttons of industrial systems such as those in electrical substations, which are high-risk and require humans to wear protective gear.

It had tested a traditional training method on the robots for two months, but they failed to perform the required tasks.

“Our collective goal is to move beyond robots that execute individual, pre-defined actions to robots that understand their immediate surroundings and instructions and can carry out complex operations,” a TetraBOT spokesman said of the collaboration with Sapiens.

Another potential customer – Shanghai Xinwei Medical Technology, a medical implants maker – is working with Sapiens to forecast inventory gaps and generate business insights.

The company struggles with using off-the-shelf AI models because they do not understand how different consumables are linked and the relationships between distributors in their procurement pipeline.

Xinwei is working with Sapiens to build custom AI models powered by industry data and a self-learning system that can adapt to information gaps.