From bits to atoms: Lessons from AI’s leap into the physical world

In the years since Grab first launched in Southeast Asia, we’ve learned to operate in the world of bits: We orchestrate millions of rides and deliveries across the region, making sense of data streams and algorithms.

Now, we’re entering a new inflection point. As AI leaves the screen and enters the physical world, Grab is evolving into a bits-to-atoms company.

It’s the start of the era of embodied AI—which is to say, AI that lives in physical hardware like autonomous vehicles (AVs), delivery robots, and other specialised robots deployed in the real world. We see robotics as very much part of the next stage of our evolution. We aim to be a global leader in urban embodied AI.

No longer just a futuristic concept

We’re already well on this journey.

Grab has an autonomous passenger shuttle service in Punggol, Singapore, under the Ai.R brand, since April 1st. You can go to rideair.ai right now and book yourself a ride there. This is the first deployment of autonomous vehicles in residential areas in Southeast Asia. It has already clocked over 40,000 kilometers and served over 3.8K public rides, without any safety incident.

In terms of autonomous delivery robots, we have just announced our collaboration with the Punggol Digital District to deploy autonomous delivery robots there. We’re making headway in terms of practical experience, putting robots out there in the field.

But let’s take a step back to think about the challenges that exist, especially for a company like Grab that previously operated in the software space.

The challenges are real and different from anything we’ve faced before. Software, and internet software in particular, exhibits the characteristics of zero marginal cost. So once you develop the software, and put it on the internet, distribution is almost free. Every incremental user you add, every incremental transaction doesn’t cost much. In contrast, every robot you build costs money. Every design decision you lock in, you’re committed to. If you need a new sensor after 4,000 units are in the field, that’s a real cost, not a simple redeployment.

There are technical challenges as well. With software, if you change your mind about a design decision, it’s easy. You can just deploy the change and roll it out to users. Contrast this with robots. Let’s say you have several hundred robots out there and you need a new set of sensors. Upgrading or exchanging them comes with real costs. 

Data is the third constraint. The internet is data-rich and allows fast feedback loops, especially at platform scale. The physical world isn’t. You have to collect it yourself, through your robots and physical sensors, which means the feedback loop is slower and smaller scale. Getting this right is what separates teams that improve quickly from teams that don’t.

Solving these challenges is a work in progress. For one, we are investing into simulations and testing, both in closed and open courses. We make sure to get it right at a small scale before we scale up. We also ensure that we’ve laid the foundations for a solid data flywheel, where we have a feedback loop between data gathered by robots in the field and the training data that continuously improves the next generation.

Beyond our own scale and abilities, we see partnerships playing a critical role, both in terms of supplementing our data collection, but also in terms of integrations. The more we operate, the more experience and data we collect, leading to smarter and safer deployment.