The Smart Bus Network


Look ma, no hands

Sometime last week, commit id 0078753328a37 wound its way through our deployment infrastructure and was deployed to our production servers and with it brought about a paradigm shift in the way we work.

Whole network managed by one human (and one bot)

We now run our whole fleet operations for 75 buses with one human and one bot.

First — a little math(s).

With about 75 buses in the system doing more than 200 trips in the day and clocking around 100,000 kilometers travelled per month we are not even scratching the tip of the iceberg when it comes to transit demand in Mumbai. Mumbai does 25 million paid trips A DAY! With an average daily throughput of 65 passengers per bus it will take


In other words, the above chart is actually just table stakes for entering this market. If you’re not able to scale your operations all you’re doing is cooking lunch for someone else. But there’s more —

Why to scale ops

We’re at the start of two golden ages in urban transportation — the age of the autonomous vehicle and the age of the bus. As our society transforms into an information society, cities with their dense agglomerations of human minds become disproportionately valuable as compared to, let’s say, the sparse settlements of suburbia or rural areas. Now unfortunately these minds are attached to great big hulking slabs of meat that need to be moved around (that like to move around, no less) and so meatspace needs to come up with an answer to this problem — and the answer of course is high throughput transit achieved through orders of magnitude increase in the speed and size of vehicles — which of course, perfectly describes autonomous buses.

Buses are social machines. Ever seen a bus network with one bus? Or a bus with one passenger? They look sad. Buses are not designed to work alone but in herds, or networks, and they love having lots of people on them. As these buses go from place to place following predictable schedules they form a network and they depend entirely on this network to tell them where to go and when. Autonomous buses cannot exist without autonomous bus networks and autonomous networks will be monitored and managed by bots. Humans will basically only be dealing with those increasingly rare cases where the bot made a bad call.

This is why it’s important to scale ops, to reduce the amount of work done by human hands in the running of buses. Our job as network designers is to enable and engender the information flows that allow an autonomous bus network to arise, to build the systems that turn this input into an output in the real world and to engineer systems that do this safely while dealing with 1000s of tonnes of flesh and blood and diesel and steel.

What is a Smart Bus Network?

In the end, human mobility is an indicator of the health of our cities. More and better mobility for more people is an unequivocally “good” thing. And although this mobility has always been expensive, we are on the cusp of a pardigm shift. 10 years from now we will wonder how we allowed so many people to die, directly and indirectly, as a result of our mobility choices. We will wonder how we made any progress as a society when our average travel speeds were no faster than those of a horse-cart (and scarcely more comfortable). And we will be aghast at the sheer waste of space, time and energy that we bore while doing so.

But mobility is hard. It is after all the cardiovascular system that powers the emergent consciousness of a city. Linking this system with the whole city in a manner that it flows naturally where it is required is an exercise in city scale self awareness that has only recently become possible. And it has, actually, become possible, to build an emergent consciousness whose role it is to provide amazing transit for humans. This is the exciting journey modern mass transit finds itself on — the creation of self adminstering systems of transit that maximise efficiency across the system.

Having said that, a smart bus network is a complex thing and is built and evolved piecemeal. Below we lay out our roadmap to achieve our vision as a series of capabilities that we’re building into the network.

Level 0 — Self awareness

  • It should know what and where its constituent parts are
  • It should know where its constituent parts should be and reliably detect anomalies
  • It should have a basic model for the geography it covers, in terms of distance, time and cost.
  • It should have a way to gauge the health of each of its constituent parts
  • It should provide forensic evidence to debug and investigate any incident on the network.

Level 1 — Self Reflection

  • It should have a model to calculate the overall reliability of the network
  • It should have a model for the financial health of the network.

Level 2 — Awareness of external world

  • It’s knowledge of geography can now be considered sophisticated, including demographic information and some prediction ability. It also has awareness of changes in geography (road blocks, accidents, new road construction, etc.)
  • It should know what people are and where and when they usually travel
  • It should understand a concept of vendor and financial and operational mechanics of the supply side.

Level 3 — Reaction to stimuli

  • All lower level intelligence should be operating at an advanced level (cameras, OBD devices, remote and machine monitoring of the fleet)
  • The system can recalculate or reroute on the basis ofperturbations such as temporary road closures, festival times, special large events (cricket matches, religious gatherings) and so on
  • The system can suggest recalibrations to financial incentives and penalties for the supply side.
  • The system can track and react to changes in customer satisfaction.
  • The system can suggest recalibrations to the fleet mix as well as marketing channel mix
  • The system can be depended upon entirely to escalate matters to human overseers.

Level 4 — Intelligence

  • The system can understand simple cause-effect mechanics
  • The system can develop a unified view of all its parts (fleet, customers, routes) and has a model of the interactions between them.
  • The system can deploy additional resources or dial down resource allocations (fleet, marketing, customer support) as and when required.
  • The system can self heal (!)

Level 5— Transcendence

The system integrates fully into all other urban smart systems and its artefacts recede into the background while still powering a mobility network that is orders of magnitude more efficient than the one it replaces today.

Where we are today

Even though the way forward seems clear now, it wasn’t always so. Formulating the model and getting to level 0 were major slogs mostly due to gaps in our world view. We hadn’t yet found the “points of balance” on which this structure would be erected. Until we got here we didn’t even know if our vision was even possible. Now we don’t expect any major conceptual barriers until we close the feedback loops at level 4 — its mostly simple execution. Better instrumentation, more bot-readable dashboards, cleaner data pipelines — all very exciting to build but no more hard ideas required. At Level 4 though we’ll find out if we’ve created a serene machine that dreams up buses or a monster whose inscrutable mind does more harm than good.

Our whole system is definitely at Level 0, with some parts at Level 1. 100% GPS coverage, basic geographical models and reasonable transparency around system reliability have been achieved. A total understanding of constituent parts, partial understanding of geography and a basic anomaly detection system allow a bot to keep escalating matters to human overseers who can make quick decisions based on thumb rules. This is where the 75/1 bus/human ratio comes in.

And we have a cool logo

So yes, bring on the robot buses! We’re ready for them.

In subsequent posts, Limo Operations team will talk about how we scaled up ops with some specifics and details. If you’re a transit nerd you should definitely follow us here.

Leave a Reply