At Vayu, we are building environmentally sustainable and scalable robotics solutions. Our first application is an autonomous robot that can do last mile delivery on public roads. Our robots are trained to operate safely in bike lanes and along the right edge of a road when the road is shared with other vehicles. They have the weight and speed of a human riding a bike, making them much safer than traditional autonomous vehicles. Since they do not occupy an entire driving lane, they don’t contribute to regular vehicular traffic. They can move faster than sidewalk robots because they drive on roads. Being lightweight, they have a much lower carbon footprint than an electric vehicle. We believe that delivery robots of this size and speed are an environmentally sustainable and economically viable solution to the last mile delivery problem.
Vayu Drive
Vayu Drive is the autonomy stack that drives our robots. It consists of differentiable yet interpretable neural network modules that are end-to-end trainable. These modules learn to represent the state of the world as well as the uncertainty in knowing this state. They predict how the state will change as a result of the robot’s actions, creating a “world model,” and learn to plan using this. We believe that these learned representations of state and uncertainty will ultimately scale better than hand-designed intermediate representations because they are limited by compute rather than the ingenuity of their designers.
Vayu Drive is trained using a combination of simulated and real data. Robust driving behaviors can be learned far more easily in simulation, where it is possible to generate adverse driving conditions and corner case scenarios, than in the real world where they are encountered only by chance. Tools from reinforcement learning, imitation learning, and adversarial self play, accelerated by generative modeling, give us the ability train robust models in simulation. However, transferring these behaviors reliably into the real world has been a well known challenge in machine learning research. We have been working on solving this sim-to-real problem and have reached a key milestone in showing our driving agent’s ability to learn how to drive in simulation environments and transfer that driving ability to the real world with a limited amount of real-world data collection. Using hundreds of hours of simulated driving and only 30 minutes of real-world data collection, we were able to train Vayu Drive to solve the lane keeping task.
Sensor configuration
Vayu Drive uses only cameras for perception. Six RGB cameras are used to get a 360-degree surround view. However, for the lane keeping task, we found that one front camera was sufficient.
Training in simulation
Our driving agent is first trained in simulation where it encounters challenging and diverse driving situations. This training happens over hundreds of simulated hours.
Real-world data collect
Real-world data is used to adapt the simulation-trained agent. The agent must drive in closed loop in simulation while simultaneously explaining real data. The goal of this training is to make the agent use the same internal world model to explain both sim and real domains.
Real-world closed loop behavior
The adapted model shows promising lane keeping behavior in bike lanes. This is remarkable given the small amount of real-world data used. We did not use HD maps, localization, or expensive sensors such as LiDAR. The autonomy stack consists almost entirely of neural networks and runs in real time on an embedded SOC platform. Very few lines of code were needed since most of the work happens in the neural network weights.
This is only the beginning
We are developing more sophisticated interactions among agents in simulation and aim to show the transfer of more complex behaviors from sim to real. We have built a training pipeline that allows us to iterate quickly and bring up new features without writing many more lines of code. Our approach is not limited to bike lane autonomous robots and can be easily extended to solve the autonomy problem in other domains including warehouses, factory floors, grocery stores, and shopping malls.
We believe that in order to solve robotics in a scalable way, it is necessary to absorb large amounts of experience. Gathering this experience in the real world can be difficult, expensive, and unsafe. Generating experience in simulation and distilling it into neural networks is much more scalable. At Vayu, we are working on finding ways to align this distilled knowledge with the real world. We are making algorithmic improvements as well as scaling our training infrastructure.
If you love robots and want to see them scale, please connect with us!
