Field guide
Last updated:
What is VLA with low-data
Why low-data vision-language-action training prioritizes rich, high-quality episodes over high-volume teleoperation when building policies for moving agents.
What is VLA with no data
How vision-language-action policies can improve with little or no new demonstration data by leaning on world models, imagination, and closed-loop reinforcement learning.
What is World Action Model (WAM)
World Action Models, championed by robotics leaders like NVIDIA’s Jim Fan, replace language-model pre-training backbones with video generation and predictive dynamics so robots can dream before they act.
Low-volume training for moving agents: Dynamic Intelligence vs Physical Intelligence, Figure AI, Skild, and Tesla Optimus
How Dynamic Intelligence’s low-volume, high-quality and no-data RL methods compare with Physical Intelligence, Figure AI, Skild, and Tesla Optimus in the race toward general-purpose intelligence for moving agents.
World models: the next state of reality
Why an LLM predicts the next token but a world model predicts what happens next in physics, and how that difference is the bridge from reacting robots to planning ones.
World models in robotics
What world models are, how they support robot training and simulation, and which ideas, from data curation to post-training, matter on the floor. Framed for robotics teams; concepts align with NVIDIA’s physical-AI glossary.