New reinforcement learning research from Physical Intelligence shows how robots can improve precise manipulation skills using minutes or hours of real-world experience.
Robot aligning a screwdriver with a tiny screw during training.
Physical Intelligence is developing foundation models designed to bring general-purpose AI into the physical world. Rather than building intelligence for a single robot or application, the company is working toward models that can control different robotic systems and perform a broad range of physical tasks.
Its models have already demonstrated tasks ranging from folding laundry and making coffee to cooking and manipulating everyday objects. As these models become more capable, however, broad knowledge is only part of the challenge. Robots also need the precision, speed, and reliability required to perform physical work consistently in real environments.
Physical Intelligence's latest reinforcement learning research explores how robots can improve specific skills through practice. Its new method, called RL Tokens (RLT), allows a pretrained vision-language-action model to adapt difficult parts of a task using relatively small amounts of real-world robot data.
Instead of retraining the entire foundation model, RLT creates a compact representation that a much smaller reinforcement learning system can use to refine the robot's existing behavior. This allows the robot to practice a difficult movement, learn from the results, and improve with minutes or hours of real-world experience.
Physical Intelligence tested RLT on four tasks where small errors can determine whether the robot succeeds: using an electric screwdriver to drive a small screw, fastening a zip tie, inserting an Ethernet cable, and plugging in a power cord. In each case, the foundation model could already perform much of the task but struggled with the final movements that required precise positioning and contact.
Rather than retraining the robot on the entire task, RLT focuses learning on these difficult moments. Physical Intelligence found that the robots could improve with as little as 15 minutes of real-world robot data, increasing both speed and success rates. Across the four experiments, the approach improved the speed of the most precise stages by as much as three times.
The larger idea behind the research is that robots could continue improving after they leave the lab. A general-purpose model could provide the robot with a broad set of capabilities, while real-world experience helps it refine the specific skills required for the environment and work it encounters.
Physical Intelligence describes this as moving toward robots that can learn directly on the job. A robot deployed in a factory, warehouse, or other workplace could encounter a difficult part of a task, practice it, receive corrections when necessary, and gradually become better at performing it. That creates a path where deployment itself becomes part of the training process rather than something that happens only after training is complete.
RLT points toward a model where robots do not need to arrive with every task perfectly learned in advance. Foundation models can provide broad physical capabilities, while smaller amounts of experience can help robots adapt those capabilities to specific tasks, equipment, and environments.
That could become increasingly important as robots move into workplaces where conditions vary from one location to another. Instead of creating a separate model for every application, developers could start with general-purpose intelligence and allow robots to develop more specialized skills through experience in the environments where they are actually used.
Real-world deployment could become an important part of how the next generation of physical AI is developed. Businesses can provide the environments, workflows, objects, and tasks that robots need to encounter, while robotics companies bring increasingly capable models and hardware into those environments.
Our Real-World Robot Training Data project is exploring ways to connect these two sides and create more opportunities for robots to train, test, and improve through practical deployments. If you are developing physical AI or operate a business interested in exploring real-world robotics, we would be interested in connecting.
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.
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.
