Roboticspost
Physical Intelligence Is Teaching Robots to Learn on the Job
New reinforcement learning research from Physical Intelligence shows how robots can improve precise manipulation skills using minutes or hours of real-world experience.
70
Reflex Robotics is deploying general-purpose humanoid robots into logistics and industrial environments, using real-world work and human supervision to help its systems become increasingly capable.
Reflex robotics deployed in real world work environments.
Reflex Robotics is developing general-purpose humanoid robots designed to perform repetitive physical work in logistics and industrial environments. Rather than waiting for humanoid robots to become fully autonomous before putting them to work, the company is taking a deployment-first approach that combines autonomous capabilities with human supervision.
The strategy allows robots to begin operating in real workplaces while the underlying technology continues to improve. By working inside existing operations, robots encounter the objects, equipment, workflows, and unexpected situations that are difficult to fully reproduce in a controlled development environment.
One of the interesting aspects of Reflex's approach is its focus on fitting robots into workplaces as they already exist. Its robots are designed to interact with equipment businesses already use, including carts, scanners, monitors, touchscreens, and other tools rather than requiring an entirely new environment built around automation.
This opens up a much broader range of potential applications. Warehouses, fulfillment centers, manufacturing facilities, and other industrial operations contain repetitive workflows that were designed around human workers. A general-purpose robot capable of using those same tools and moving through those same environments could allow businesses to explore automation without redesigning every part of their operation.
Real workplaces contain situations that robots may not yet know how to handle on their own. Reflex addresses this with a supervised reliability system that allows human operators to step in when a robot encounters a difficult or unfamiliar situation, helping it continue working instead of bringing the entire deployment to a stop.
This approach allows useful deployments to begin before every possible edge case has been solved through autonomy. The robot can handle the portions of a workflow it understands while human supervision provides support when necessary, creating a practical bridge between today's capabilities and increasingly autonomous systems.
Every deployment also creates opportunities for Reflex's systems to improve. As robots perform repetitive workflows and encounter new situations, supervised interactions provide examples of how those situations should be handled in the future.
Over time, this creates a feedback loop between deployment and development. Robots can be introduced into real operations, learn from the work they encounter, and become more capable as experience accumulates across deployments. For physical AI, the workplace becomes more than the final destination for a finished robot. It can become part of the learning process itself.
Logistics provides a strong environment for this deployment model because warehouses and fulfillment centers contain large numbers of repetitive physical workflows. Reflex has been working with GXO Logistics to deploy its humanoid robots in warehouse operations, where the companies are developing use cases around tasks that would otherwise require repetitive manual work.
The partnership is structured around a Robots-as-a-Service model, giving GXO a way to introduce the technology into its operations without treating the robots as a traditional equipment purchase. More importantly, deployments like these give Reflex an opportunity to develop its systems around the requirements of a real logistics operation and learn from the conditions robots encounter while working.
Reflex's approach demonstrates why access to real businesses could become increasingly important for robotics development. Warehouses, factories, machine shops, hotels, and other workplaces contain different workflows and environments where emerging robotic systems can be tested, trained, and eventually put to productive use.
Our Real-World Robot Training Data project is exploring ways to connect robotics companies with businesses interested in participating in these kinds of deployments. If you are developing robotics technology or operate a business with repetitive physical workflows that could provide an interesting deployment environment, we would be interested in connecting.
New reinforcement learning research from Physical Intelligence shows how robots can improve precise manipulation skills using minutes or hours of real-world experience.
Robots are getting more capable, but the next challenge is giving them enough experience in the environments where they will actually work.
Creating real-world deployment opportunities where robots can perform practical tasks and generate valuable training data for physical AI.
A fully open-source AI robot featuring natural conversation, autonomous navigation, computer vision, and a modular 3D-printed body.