BREAKING
NVIDIA Unveils World Action Models
From VLA to WAM
1
Pretrain video dynamics
↓
2
Inherit physics priors
↓
3
Generate robot actions
0
B
parameters
0
Hz
control rate
WAM vs VLA Task Progress
WAM
62.2
VLA
27.4
Unseen tasks
39.5
Strengths and Open Questions
Strengths
●
30 min, 55 trajectories to adapt
●
Human-to-robot data efficiency
Challenges
●
High compute load
●
Monocular spatial accuracy
Open-Sourced for Developers
AI NEWS BLITZ
NVIDIA has introduced World Action Models, a new approach to controlling robots in the real world.