NVIDIA's AI Learns Why Copying Humans Isn't Enough
NVIDIA combines imitation learning and goal-based RL to train parkour AI on just 30 seconds of data
“Systems that copy us are beautiful, but brittle, repetitive. Systems that chase goals adapt, but often stop moving like us.”
NVIDIA researchers developed a hybrid approach that trains virtual humans to perform parkour by combining motion imitation with goal-directed reinforcement learning, using only 30 seconds of internet video as training data. The system generalizes to unseen obstacle arrangements and composes novel movement sequences, overcoming the brittleness of pure imitation learning. While technically impressive, this is a simulation/animation research result with limited direct AI industry signal for enterprise or LLM-focused readers.