Predicting physical trajectory outcomes through video forecasting before motors actuate.
Robots in homes cannot rely on brittle pre-programmed motion scripts. They require real-time physical simulation paired with continuous conversational understanding.
The 1X World Model is a high-resolution generative video model trained on millions of hours of real-world physical interactions. Before NEO moves its arm, it simulates the outcome of each potential action across time, selecting the trajectory with the highest evaluated probability of success.
Redwood AI is 1X’s generalist end-to-end model. It seamlessly fuses conversational voice understanding, high-rate visual perception, and coordinated whole-body motor control. Redwood learns continuously from both successful chore completions and unexpected physical failures.
Trained using massively parallel physics simulation and transferred directly to hardware via zero-shot sim-to-real policies. NEO navigates carpets, hardwood, uneven clutter, and staircases with compliant knee and ankle reflex loops.
From video trajectory prediction to whole-body coordinated manipulation in living rooms.
The 1X World Model now serves as NEO's primary cognitive evaluation layer. Ask NEO to clear an unfamiliar table setting, and watch it reason through spatial constraints without pre-programmed heuristics.
Published experimental findings comparing predicted trajectory outcomes against real-world execution. The model achieves 94.8% accuracy in forecasting task success prior to motor actuation.
Redwood AI's latest reinforcement learning (RL) policy unlocks full bipedal agility—fluid walking, kneeling, climbing household stairs, and soft compliant ground recovery.
Redwood unifies bipedal balance with bilateral manipulation. NEO leans its torso and braces with one hand while reaching deeply into low cupboards with the other.
We are actively hiring researchers and systems engineers to scale physical foundation models.