ROBOTS
Simulation-trained walkers
Humanoid walkers trained entirely in simulation, shown with the conditions they were tested under.
Obstacle Course
A humanoid walker trained in simulation using only joint and inertial sensing, replayed from its recorded simulation states on obstacle fields it never trained on. 1 of 24 attempts lasted the full 40 seconds; it is shown first, at 2.5× speed. Then a 39-second, 12-meter attempt that ends in a fall, in real time. The other 22 attempts fell sooner. The ground is colored by height for viewers only.
The Raccoon and the Robot
A simulation-trained walker follows a raccoon through a generated village. In this scene only the robot's feet collide with the world; fences and buildings are scenery. It ends in a fall. Development footage.
What these clips are
Both clips replay recorded simulation states; nothing is animated by hand. The obstacle-course walker uses only sensors the real robot has: joints and inertial measurement. The village walker is an earlier controller that also reads its own body velocity.