Robotics and Embodied AI

This page tracks robotics, embodied intelligence, morphogenesis/biology-adjacent AI ideas, and long-horizon physical control.

Sources in this batch

• Michael Levin’s biology-oriented video suggests ideas from biological organization may inform AI systems.
• Figure’s HQ tour provides context on humanoid robot development.
• Physical Intelligence’s robot Olympics post invokes Moravec’s Paradox.
• DexDrummer demonstrates in-hand, contact-rich, long-horizon dexterous robot drumming.

Research interest

The surprising through-line is that embodiment exposes the weakness of text-only or benchmark-only intelligence claims. DexDrummer and robot Olympics-style tasks are interesting because contact-rich manipulation, long horizons, and real-time control are hard to fake with next-token imitation. Levin-style biological framing is worth tracking if it yields computational mechanisms rather than metaphor.

Open questions:

• Which foundation-model techniques transfer to contact-rich robotics?
• Can morphology, self-organization, or developmental biology inspire robust control algorithms?
• What benchmarks capture the long tail of physical-world failure?

Related

• world-models-and-video-intelligence
• ai-for-scientific-discovery
• llm-evaluation-and-belief-management

Remainder-batch update

New related sources include Tesla FSD competition in China, Figure 03, robotic industrial design material, Unitree security concerns, and Simone Giertz’s older breakfast-machine example. These reinforce the page’s focus on physical-world deployment and embodied-system reliability.

Updated: 2026-06-27