As generative AI models transform how humans interact with software, a parallel race is unfolding in the physical world: the pursuit of General Physical AI. But unlike language models, which learn from the vast and readily available corpus of internet data, robots face a more constrained and costly landscape. Physical intelligence requires more than data—it requires experience. And experience in the real world is expensive.
Robots cannot scrape the internet for real-world motion. There’s no YouTube for sidewalk navigation or door-to-door delivery. Autonomous vehicles benefit from millions of human drivers acting as implicit teachers, but most robots lack a comparable population. The road to General Physical AI will be paved not with code alone, but with deployment: millions of robots operating in the wild, generating the raw experience needed to train deep neural networks for the real world.
At RIVR, we are pioneering this future by deploying robots in last-mile delivery—giving one human the power of 1,000 robots. Our approach fuses artificial neural networks with a breakthrough wheeled-leg design, enabling robots to move fluidly through complex urban terrain. With a mission to deploy over a million robots, we’re laying the groundwork for a powerful data flywheel—one that accelerates robotic intelligence with every delivery made.
The Bet on Delivery: A Gateway to 1M+ Robots
RIVR believes the answer lies in delivery. Specifically, last-mile and last-100-yard delivery. With over 166 billion packages delivered annually around the globe, the scale is already there. Add in food and grocery delivery, and it is expected to be an addressable market worth $1-2 trillion. More importantly, this is a domain rich in general robotic challenges: mobility, manipulation, human interaction, and unstructured environments.
Each delivery is a training repetition. A chance for the robot to navigate new terrain, interpret visual scenes, avoid pedestrians, open gates, and interact with humans. RIVR’s strategy is to use this repetition as a source of experience—a way to build the foundational knowledge that will power General Physical AI.
Vertical Focus, General Outcome
While some pursue a horizontal approach to Physical AI—aiming to build general software platforms for any robot—RIVR is taking a vertical-first path. This focus allows for rapid iteration, tight integration between hardware and software, and meaningful real-world deployment.
Why not horizontal? The answer lies in the scarcity of meaningful data. OpenAI built ChatGPT on the back of a massive internet corpora. Robotics lacks this luxury. Horizontal models only work when there is enough shared data to justify abstraction. Until such data exists, the only viable path is to go vertical: find a high-volume, high-complexity use case, and own it end-to-end.
Delivery meets that bar. It requires solving urban autonomy, social navigation, last-yard manipulation, and real-time decision-making. Mastering delivery creates a playbook for more general use cases. In other words, RIVR’s focus on delivery is not a constraint—it’s a launchpad.
The Deep Learning Breakthrough
We integrate both Reinforcement Learning (RL) and Supervised Learning (SL) into our neural networks, developing a robust Physical AI that mimics human learning—combining imitation with trial-and-error learning. This integrated strategy allows us to handle robotic challenges through a unified neural network architecture, streamlining software coordination and reducing computational demands.
RL with Simulated Data: The Foundation of Rapid Learning
Our approach to RL is built on large-scale simulation. Using parallel GPU simulators like NVIDIA’s Isaac Sim, we train legged robots through trial-and-error learning, dramatically reducing training time. This method allows us to:
- Simulate complex real-world physics, including electric actuators and sensor models.
- Develop proprietary learning algorithms that optimize robot behavior.
- Create game-like training worlds that refine mobility, manipulation and autonomy.
By the time our robots are deployed, they have already encountered and adapted to millions of real-world scenarios in simulation, allowing them to master new tasks in just days.
SL with Real-World Data: The Data Flywheel
To complement RL, we leverage SL with real-world data, similar to techniques used in autonomous driving. By collecting expert data in the field, our robots continuously refine their understanding of complex urban environments. This data flywheel enables each deployed robot to gather more insights, enhancing autonomy and accelerating deployment. As more robots enter the field, they generate even richer data, ultimately solving the long tail of autonomy.
The Endgame: General Physical AI That Coexists with Humans
The ultimate goal is a world where robots can navigate sidewalks, open doors, hand off packages, and communicate with people as fluidly as human couriers. Delivery is simply the proving ground. The repetition. The training camp for General Physical AI.
By mastering the last 100 yards, RIVR is laying the foundation for robotic intelligence that scales beyond logistics. Into all urban mobility. The next frontier of AI isn’t just about what machines can say—it’s about what they can do. And for that, they need to learn from the world. One delivery at a time.
.png)
