NVIDIA-backed team builds parkour-capable humanoid robot
Humanoid robot learns parkour skills without hand-coded rules or motion graphs
Deep Dive
Researchers present Light-Loco-Parkour (LLP), a perceptive whole-body locomotion system for humanoid robots. Driven by a single reinforcement-learned policy that uses only onboard depth and a velocity command, it autonomously decides when to walk, balance, climb, step down, or vault—without reference input, skill labels, or hand-coded state machines. The same policy transfers zero-shot to real-world indoor and outdoor hardware experiments.
Key Points
- LLP is a single policy that handles walking, climbing, vaulting, and other whole-body locomotion skills without motion graphs or hand-coded rules
- Achieves zero-shot transfer to real-world hardware using only onboard depth sensors and velocity commands
- Developed by NVIDIA and Tsinghua University researchers, with code available on Hugging Face and DagsHub
Why It Matters
Pioneers autonomous humanoid robots that can navigate real-world obstacles without pre-programmed behavior trees or motion libraries.