NVIDIA's Jim Fan Unveils Physical AGI Roadmap: WAM and DreamDojo
NVIDIA researcher Jim Fan outlined a roadmap toward physical AGI at Sequoia AI Ascent, arguing Vision-Language-Action models fall short and proposing World Action Models (WAM) as the second pretraining paradigm for robotics — mirroring the LLM success story. Supporting work: EgoScale (a Dexterity Scaling Law for robot manipulation data) and DreamDojo, an end-to-end neural physics engine for scaling reinforcement learning in silico to close the sim-to-real gap.
Why It Matters
Fan's framing parallels the LLM playbook — massive pretraining, scaling laws, data flywheels — applied to physical AI. If validated, WAM and DreamDojo could do for robotics what transformers did for language: collapse the timeline to capable, general-purpose physical agents.