Air Street Press Embodied AI is hitting its stride
Dec 9, 2025
Explore the robotics renaissance driven by foundation models, revealing how neglected fields are rapidly advancing. Discover innovations in world models like Odyssey 2 and Dreamer 4, showcasing long-horizon capabilities. Delve into the unification of perception and action with VLAMs and the latest humanoid control breakthroughs. Examine planning layers that enhance interpretability and execution, along with real-world deployments like Seriact Cortex in warehouses. Finally, witness large-scale driving with Wave's models in 90 cities, pushing the limits of embodied AI.
AI Snips
Chapters
Transcript
Episode notes
World Models Become Rich Training Grounds
- World models now support interactive, agent-conditioned rollouts and 3D persistence enabling richer training grounds.
- These models can become cheap, safe substrates to stress-test embodied policies before real-world exposure.
VLAMs Replace Custom Perception Pipelines
- Vision-language-action models (VLAMs) unify perception and control under shared multimodal representations.
- VLAMs outperform modular pipelines on generalization, long-horizon tasks, and cross-embodiment transfer.
Planning Layers Improve Interpretability
- Explicit planning layers produce mid-level plan tokens or spatial anchors that decompose tasks before low-level control.
- These planning tokens improve interpretability, controllability, and enable cross-embodiment transfer.