Convergence

UPDATED · APR 22 2026

Connections between concepts across different KB topics. These represent convergence points where breakthroughs in one field directly enable progress in another.

Domain pairs

Connections

Live cross-links

AI ↔ Robotics

AI Concept	Robotics Concept	Connection
Agentic Reasoning	Foundation Models for Robotics	Three-layer agentic framework (foundational → self-evolving → multi-agent) maps to robot skill levels. GPT-4V as the reasoning layer in Humanoid-COA is a direct instantiation of agentic reasoning in physical systems.
LLM Tool Use	Humanoid Loco-Manipulation	Robot actuators are "tools" in the agentic framework. The three paradigms (prompting → SFT → RL) predict the evolution of robot control from zero-shot to fine-tuned to RL-optimized.
Reinforcement Learning for Agents	Sim-to-Real Transfer	RL is the training paradigm for sim-to-real. Credit assignment in long tool chains (AI problem) = credit assignment in long manipulation sequences (robotics problem).
Evolutionary Algorithm Discovery	World Models	AlphaEvolve's evolutionary search could optimize world model architectures. Automated discovery of better simulation parameters.
World Models	World Models	Dual-citizen concept. V-JEPA 2-AC (Meta FAIR) achieves zero-shot Franka manipulation after <62h of robot data on top of 1M+ hours of passive video pre-training — the strongest current evidence that AI-side world-model research translates to robotic control. H-WM and StructVLA further operationalize JEPA-style abstract prediction for TAMP and manipulation. The industrial track (1X NEO, Isaac Sim) and research track (V-JEPA 2, H-WM) are converging.
JEPA	Sim-to-Real Transfer	JEPA's abstract-representation prediction is architecturally complementary to sim-to-real: JEPA learns what to pay attention to; sim-to-real handles physical realism gaps. V-JEPA 2-AC deployed zero-shot without domain randomization tricks.
Generative World Models	Sim-to-Real Transfer	Wayve GAIA-2's pixel-space generative world model is used for sim-to-real training data augmentation at a production AV company. Generative models' physics-consistency failures (PhyWorldBench: ~60%) are tolerable for augmentation, catastrophic for control — explains the camp specialization.

AI ↔ Optical Computing

AI Concept	Optical Concept	Connection
Agentic Reasoning	Photonic Neural Networks	AI model inference at scale requires photonic acceleration. Agent systems that make thousands of inference calls per task need the sub-nanosecond latency and femtojoule efficiency of photonic computing.
AlphaEvolve	Photonic Tensor Cores	Evolutionary algorithm discovery (AlphaEvolve) could design optimized photonic circuit layouts. The 32.5% FlashAttention speedup demonstrates the pattern.

AI ↔ BCI/Neuroscience

AI Concept	BCI Concept	Connection
Chain-of-Thought Reasoning	Neural Signal Decoding	Deep learning advances that improve LLM reasoning also improve neural signal decoding. Same architectures (transformers, attention) applied to both domains.
Mechanistic Interpretability	Invasive vs Non-Invasive BCI	Understanding how AI models process information (mech interp) parallels understanding how brains process information (neuroscience). Circuit tracing in LLMs mirrors connectomics.

Robotics ↔ Electrification

Robotics Concept	Electrification Concept	Connection
1X Technologies	Sodium-Ion Batteries	Consumer humanoids ($20K NEO) need cheap, safe batteries. Sodium-ion's cost advantage could enable the economics of household robots. Battery weight/energy density directly constrains robot operating time.

Optical Computing ↔ AI Infrastructure

Optical Concept	AI Connection	Impact
Lightmatter	Google DeepMind	Lightmatter's 200 Tbps/package interconnect directly enables the scale of AI training that produces models like Gemini. AlphaEvolve's 1% training speedup could compound with photonic interconnect improvements.
Photonic Interconnects	AI data center infrastructure	The bandwidth bottleneck (GPU-to-GPU communication) that photonic interconnects solve is the same bottleneck that limits AI model scale. 8x faster training from L200 = larger models = better agentic reasoning.

Space ↔ Robotics

Space Concept	Robotics Concept	Connection
On-Orbit Servicing	Sim-to-Real Transfer	Satellite refueling requires robotic manipulation in zero-g. Sim-to-real is critical because you can't debug in orbit. The 99% sim-to-real correlation has direct implications for space robotics.

Method

Convergence points are where a breakthrough in one domain directly enables progress in another. Linked concepts open their Knowledge Base pages; the rest are tracked but not yet written up. Back to the Atlas →