Loading
Loading
The open questions that smart people disagree on. Where you land shapes what you build.
Humanoid
Human environments are designed for human bodies. Stairs, doors, handles, tools — all optimized for our form factor. A humanoid robot can operate in any human workspace without modification. One platform, infinite use cases.
Specialized
A vacuum cleaner doesn't need legs. A warehouse picker doesn't need a face. Purpose-built robots outperform general-purpose ones in every specific task. The Swiss Army knife loses to the chef's knife in the kitchen.
Our Read
Both win in different contexts. Specialized robots dominate near-term (Roomba, warehouse bots, surgical arms). Humanoids win long-term IF the cost drops below the value of not having to redesign environments. The real question is whether general-purpose intelligence makes the humanoid form factor viable before specialized robots saturate their niches.
End-to-End
Train a single neural network from pixels to motor commands. No hand-crafted perception, no separate planner, no control module. The model learns what matters. This is how VLAs work — and it scales with data and compute.
Modular
Decompose into perception → planning → control. Each module can be tested, debugged, and improved independently. When the robot fails, you can pinpoint which module broke. End-to-end is a black box — when it fails, you have no idea why.
Our Read
End-to-end is winning in research (RT-2, π₀). Modular still dominates in production deployments where explainability and safety certification matter. The likely convergence: end-to-end for manipulation tasks, modular for navigation and safety-critical functions. Hybrid architectures that blend both will probably dominate.
Teleoperation
Humans remotely control robots to generate training data. High-quality demonstrations, clear task completion, safe exploration. This is how most robot foundation models are trained today. Scale by hiring operators.
Autonomous
Sim-to-real transfer + reinforcement learning. Train millions of episodes in simulation, transfer to reality. Scales infinitely without human labor. This is how Waymo and NVIDIA Isaac Lab work.
Our Read
Currently complementary, not competing. Teleoperation provides the seed data for initial capability. Sim-to-real provides the scale. The gap: sim-to-real transfer still loses 35-40% performance crossing to reality (95% sim → 60% real). Whoever closes this gap fastest wins — it's the most important technical race in physical AI.
Domestic
Largest TAM by far. Every household is a potential customer. Consumer margins can be higher. The Roomba proved people will buy home robots. First to crack household wins the biggest prize.
Industrial
Clear ROI (replace $15-25/hr labor). Controlled environment. Corporate buyer with procurement budget. Lower safety bar (no children or pets). This is where revenue actually exists today.
Our Read
Industrial first is the consensus — and it's correct. No regulatory framework exists for autonomous robots in homes with children. The liability model (who's responsible when the robot drops a baby?) is undefined. The Roomba succeeded because its failure modes are trivial (stuck under a couch). A humanoid's failure modes are categorically different. Factory → warehouse → commercial → household.
Real
Unitree ships at $16K. Chinese vertical integration (designing + manufacturing actuators in-house) creates 50-70% cost advantage. Government subsidies, massive engineering talent pool, tolerance for lower margins. Western companies cannot compete on cost.
Fragile
Low cost may mean low reliability. Production robots need 10,000+ hour MTBF. Quality gaps show up at scale, not in demos. Also: export controls on advanced AI chips limit Chinese compute capability for on-robot intelligence. Geopolitical risk for Western buyers.
Our Read
The cost advantage is real in hardware manufacturing. The question is whether it extends to the AI layer (foundation models, fleet learning, sim-to-real). If the value concentrates in software/AI (like smartphones), then hardware cost leadership matters less. If it concentrates in integrated hardware-software (like Apple), then China wins.