BREAKING
Humanoid Walks From Raw Depth Images
Why Replace LiDAR Maps
LiDAR Height Maps
Old way
●
Odometry drift
●
Low update, latency
●
Fault misalignment
Raw Depth Vision
New way
●
Richer information
●
Direct stereo input
●
Single unified policy
Vision-Aware Behavior Distillation
1
Teacher: ideal height map
↓
2
Distill behavior
↓
3
Student: raw depth
↓
4
Multi-critic learning
One Policy, Many Terrains
Extreme Terrain
●
High platforms
●
Wide gaps
●
Stair climbing
●
Rubble fields
Fine Motion
●
Stepping stones
●
Balance recovery
●
Indoor parkour
Accepted to RSS 2026
Research Stage, Code on GitHub
AI NEWS BLITZ
A new humanoid locomotion method learns to walk directly from raw depth camera input.