We Taught a World Model to Control Robots, Drive Cars, and Pilot Drones
Odyssey-3 is a foundation world model that can power robots, drive cars, pilot drones, train AIs, and even play video games
Oliver Cameron
September 15th, 2026
We began working on world models in 2023 with the theory that they would become foundational to a new generation of intelligent systems. In 2026, we’re starting to see powerful examples of that theory becoming reality. Today, we’re sharing an early glimpse of Odyssey-3, our most powerful, foundational world model yet. Odyssey-3 is capable of powering humanoids, driving cars, controlling robots, piloting drones, training AIs, and even playing video games.
Trained on a vast collection of visual observations of the world, Odyssey-3 has developed a deep, learned understanding of physics, dynamics, cause-and-effect, human behaviors, and other concepts that make up our reality. Odyssey-3 can then draw on that learned world knowledge to solve physical and virtual tasks, with only a small amount of experiential data required per task. We call Odyssey-3 a physics agent, a new kind of agent that speaks the language of the world—not humans—enabling it to interface natively with the physical world.
A Physics Agent Speaks the Language of the World
Odyssey-3 demonstrates encouraging performance controlling physical systems with only a few hours of experiential data, drawing on the broad world knowledge it has acquired during pretraining. This experiential data pairs the physical system’s observations with the actions taken to perform a task, providing examples of how its controls are used. An action head—a learned output component attached to the world model—is trained on these examples to translate Odyssey-3’s internal representations into the actions required by the physical system.
Odyssey-3 Is the Foundation for Flexion’s Humanoid Autonomy
Today, we’re announcing a close research collaboration with Flexion, a leader in general-purpose robot intelligence with deep expertise in reinforcement learning and whole-body control. Flexion is pursuing one of the most ambitious problems in robotics: building general-purpose intelligence that allows robots to perform useful work within the buildings, tools, and environments humans already use. Their research brings together perception, manipulation, and whole-body control to develop robots that can carry out extended tasks and recover when things go wrong.
Building on Odyssey-3 as a base model, Flexion has carried out substantial research and engineering to develop the humanoid control policies demonstrated here. With only tens of hours of humanoid teleoperation data, the resulting system can perform tasks in real time, applying Odyssey-3’s pretrained representations through Flexion’s work on robot learning and control. In our evaluations, these policies generalize better to environmental changes than the VLA baselines we tested, continuing to execute tasks under lighting changes that cause baseline policies to fail. We see these results as an encouraging demonstration of what Flexion’s robotics expertise can achieve with a foundation world model, and we’re excited to continue this research together.
Controlling Humanoid
Controlling Humanoid
Controlling Humanoid
Controlling Humanoid
Odyssey-3 serving as a base model for Flexion's robot intelligence
Odyssey-3 Can Drive Cars
With only 20 hours of simulated driving data, Odyssey-3 is capable of autonomously driving a car in closed loop on streets in India, generating driving trajectories in real time using the same underlying world model that powers our robot policies. The pretrained world model remains frozen throughout training, extracting visual representations that feed into a relatively small driving policy, which predicts waypoints ahead of the car. The driving policy therefore learns to draw on an existing understanding of the world, with the simulated experience teaching it how to apply that knowledge to driving.
Driving Autonomously
Driving Autonomously
Driving Autonomously
Driving Autonomously
Driving Autonomously
Driving Autonomously
Odyssey-3 driving a car in closed loop on streets in India
We compared policies trained entirely in simulation with policies trained on real driving footage, evaluating both on busy roads with frequent distractions. Both navigated bends while maintaining their lane, handled vehicles overtaking on either side, and turned at busy junctions. On real roads, the policies trained entirely in simulation traveled about 77% as far between safety-driver interventions as those trained on real footage. Transfer in the opposite direction was substantially weaker: when evaluated in simulation, the policies trained on real footage traveled only 15% as far between interventions as those trained in simulation. This asymmetry was one of the more surprising observations, with simulated experience transferring to the road much more effectively than road experience transferred back into simulation.
Odyssey-3 Can Power Robot Arms
With only tens of hours of robot demonstrations, Odyssey-3 can learn to control a variety of robot arms and complete complex tasks, drawing on the broad world knowledge acquired during pretraining. In our experiments, we observe recovery behaviors that are absent from the training demonstrations, including reorienting a gripper after a missed grasp and retrieving an object dropped in an unusual position or orientation. These behaviors suggest that Odyssey-3’s learned physical understanding can help a robot respond to situations beyond those explicitly demonstrated. This is encouraging for sample-efficient robotics, where collecting demonstrations of every possible failure and recovery quickly becomes impractical.
Controlling Robot
Controlling Robot
Controlling Robot
Controlling Robot
Controlling Robot
Controlling Robot
Odyssey-3 driving a car in India
An important next question is how consistently these capabilities hold up across different robots and environments. To this end, we’re excited to announce a close research collaboration with Poke & Wiggle, a leader in robot data, robot policy analysis, and large scale benchmarking. Together, we’re evaluating Odyssey-3 across different bodies, viewpoints, and controls to understand where its knowledge transfers, where it breaks down, and how those findings can guide further training.
Odyssey-3 Can Pilot Drones
Following the same training recipe as our driving experiments, we train an aerial navigation policy to generate flight waypoints from recent camera observations, the drone’s motion state, and a high-level navigation prompt. The policy draws on Odyssey-3’s pretrained visual representations, with simulated flight demonstrations teaching an action expert how to translate those representations into movement. Using tens of hours of simulated drone data, we trained a policy that demonstrated stable flight in a simulated indoor setting.
To investigate what the backbone already understood before policy training, we also visualized its predictions for aerial navigation tasks with its weights frozen. In these qualitative rollouts, we observed plausible directional flight and motion around obstacles, alongside responses consistent with the world’s semantic content. These observations suggest that pretraining provides useful knowledge of spatial structure and motion.
Piloting Drone
Piloting Drone
Piloting Drone
Piloting Drone
Odyssey-3 driving a car in India
Odyssey-3 Can Play Video Games
Odyssey-3 can also provide a foundation for policies that play video games. We train these policies on gameplay recordings paired with keyboard and mouse inputs, keeping the pretrained world model frozen. During play, the policy observes recent game frames, generates controls, and uses the resulting observations to choose its next actions. The same approach used to control physical systems becomes a way to interact with virtual worlds.
Our experiments have produced extended gameplay sessions in GTA V, alongside selected examples of driving, shooting, and hand-to-hand combat. We are also seeing early evidence of transfer: GTA-trained policies have produced movement in Red Dead Redemption 2 and motorcycle riding in Sleeping Dogs without additional policy training on either title. In one experiment, a mobility policy trained on approximately two hours of GTA footage produced horseback movement in Red Dead Redemption 2, applying controls learned in one game to a different character, vehicle, and environment.
Playing Game
Playing Game
Playing Game
Playing Game
Playing Game
Playing Game
Odyssey-3 driving key commands into games in real-time
We also explored whether a driving policy trained in CARLA, an autonomous-driving simulator, could apply what it had learned to GTA V. Without additional policy training, it demonstrated road-following behavior in a different visual world with different vehicle dynamics. In one exploratory run, the car traveled nearly 300 meters over a minute of active game time. This result offers further encouraging evidence that learned driving behavior can transfer beyond the environment in which it was trained.
Together with our results in humanoids, robotics, driving, and drones, these gaming experiments suggest that Odyssey-3 is a strong foundation for generalist embodied agents. A shared pretrained world model can support policies across different bodies, environments, and controls, with early evidence that learned behaviors can transfer between them.
Odyssey-3 Is for AIs and Humans Too
Odyssey-3 can generate worlds that AIs and humans can inhabit, and where they can take actions and experience their consequences as the simulation unfolds.
For AI agents, these worlds provide environments for learning through experience, while also giving us a way to study what those intelligences are capable of and how they behave when given the freedom to act. As language models and other intelligences take on greater responsibility, we need places where we can discover potentially dangerous behaviors and understand their consequences without learning these lessons at the expense of people in the physical world.
For humans, these same worlds enable interactive experiences that evolve in response to their decisions, creating possibilities for learning, experimentation, and exploration. We believe the same underlying world model can power physical systems and generate the environments in which their intelligence can be developed and evaluated safely.
Generating Environment
Generating Environment
Generating Environment
Generating Environment
Odyssey-3 generating interactive environments in real-time
An Early Glimpse of Physics Agents
Agents in our reality, like humans, develop an understanding of the world and how it works through observation and interaction, allowing them to draw on prior knowledge when learning a new task. We can learn to operate dangerous machinery at age 18—despite never having used it before—because we have already gained an intuition for navigating space, watched countless objects in motion, seen how they respond to force, observed humans using tools in many contexts, and noted that collisions are dangerous.
Today, many robotic systems are trained on repeated demonstrations of specific tasks to acquire the ability to manipulate objects, concentrating large amounts of narrow experience to accomplish a single task. We see this as a sign that these systems are brute-forcing the problem, compensating for a lack of general world understanding with ever-larger quantities of task-specific data. Humans do not need to sit and watch thousands of hours of specific tasks before becoming capable, because so much of the necessary knowledge has already been acquired through our experience of the broader world.
Similarly, a true physics agent should possess a superhuman understanding of the world, including the physics, dynamics, and cause-and-effect that govern how our reality evolves. That understanding would give it a native interface to physical systems, enabling it to adapt to new tasks with the amount of experiential training a human needs, or less. Odyssey-3 is an early physics agent and represents a leap in the performance and maturity of world models, capable of applying its learned understanding to robots, humanoids, cars, drones, and video games with only a few hours of experiential data.
As physics agents develop, we believe they will enable us to automate increasingly complex physical work, taking on jobs that put people at risk and helping with everyday tasks that age, illness, or disability can make difficult, allowing more people to live independently. These same capabilities could accelerate experimentation in science and engineering, give us more time to spend with one another and on work we find meaningful, and eventually allow us to operate in environments we humans have had no lived experience in, including worlds beyond our own.
These same models will also change how intelligence itself develops, generating the worlds in which future agents learn, and learning themselves from what those agents discover when acting in the physical world. More capable world models will support more capable agents, whose experiences and discoveries would in turn improve the worlds that shaped them, allowing both to grow beyond the limits of the environments we can construct by hand.
We see Odyssey-3 as a big step toward this future, and we’re preparing to release it publicly in the coming weeks, with the model already in use by leaders in robotics, autonomous driving, gaming, and defense.
