How Robot Gyms Are Preparing Humanoids for Real-World Tasks

A humanoid robot may simply not be ready to be put to work in a warehouse, provide assistance at a hospital or find its way around a home it does not know until it learns something humans often take for granted: how to cope with a physical world that is anything but predictable. To meet that sort of challenge, the robotics industry is investing into “robot gyms”, purpose-built spaces for testing and training machines and subjecting them to an array of tasks.

According to a report from the Financial Times, AI startups and robotics firms are in the process of amassing real-world data and putting training facilities in place so as to make their intelligent machines have more capabilities.

The idea is relatively simple: language AI has no trouble learning from the text that is to be found in abundance on the internet, but for a robot, one must have data on the physical environment and objects, movement, and the outcome of an action whether it is a success or not.

It’s just simply not enough for the robots to learn those skills just through watching videos or going through some programmed routine. At some point in time, a robot will have to show it can put its learning to use in uncharted territory.

Figure AI, the humanoid developer, has an approach of that sort with its Index project; the company is making use of a substantial dataset of human behaviour drawn from video for this purpose.

In August, Figure stated that its data-collection platform has seen the uploading of some 16 million videos from contributors, and there are more than 44,000 active users on a weekly basis.

When it comes to the financials, the company reported $15 million in payments to those contributors and has laid out plans to put over $1 billion into computing and data resources in the next year.

Figure has also been testing whether that training can transfer into environments that are unfamiliar.

According to a September report that was published by the company, its Helix 2.5 AI system was put to the test across 30 previously unseen homes with a trio of household tasks. Figure attributes a jump in the zero-shot success rate for such tasks, up to 56% from 9%, to pretraining on human behaviour.

Those results come from the company itself and simply don’t prove that humanoid robots are ready to work in homes without anyone keeping an eye on them.

Governments are now also beginning to invest in dedicated facilities to get these robots trained properly.

With a view to creating machines that can be put to use in the public safety sphere, the Home Team Science and Technology Agency of Singapore has opened a new humanoid robotics center on September 30.

The facility’s main aim is to train robots for routine, physically demanding or dangerous work, and it also says the first real-world uses may come as early as 2028.

Despite the growing investment, there’s no doubt that there are still some major challenges that are being faced.

A robot could possibly learn a task and master it in one environment, but then may struggle if the layout changes, if an obstacle appears, or if an unexpected scenario interrupts its routine.

Developers might well find the next advance in having robots learn to adapt instead of putting on a show with some impressive feat; and robot gyms are one answer to that, a means of closing the distance between a controlled demonstration and a machine you can have at work in the ordinary world.