Source: synthesis from VC Landscape & News + Defense Tech + Dual Use libraries
Across the readings, the “future of robotics” shows up as physical AI: a race to build data + models + deployment flywheels that compound capability.
1) Physical AI framing (foundation-model shift)
- Robotics is increasingly about foundation models + data, not just mechatronics.[1][2]
- The field looks like a “GPT‑2.5 moment”: real demos, but not generalized deployment.[3]
2) Data is the moat (and it’s expensive)
- Proprietary training data is a major competitive advantage in physical AI.[1]
- Robotics data is scarce vs internet text/video, so capital + data pipelines become structural advantages.[3]
- Teleop + simulation (Sim2Real/Real2Sim) + outsourced collection are converging into “data flywheels.”[2]
3) The last mile is reliability + physics/force
- Humanoids still struggle with mundane tasks reliably (doors, stairs, “small stuff”).[4]
- Force/physics interaction and tactile sensing look like core blockers, not just planning.[4]
- Example micro-unlock: tactile + soft grippers for deformables (cables/ropes/cloth).[5]
4) Near-term value accrues to vertical/full-stack deployments
- Because models aren’t yet plug-and-play, vertical/full-stack deployment captures value first.[3]
- vRaaS is a pragmatic commercialization path: constrained tasks → uptime → workflow integration → proprietary data.[2]
5) Defense/dual-use may pull the timeline forward
- Defense is positioned as an early scaling customer (budget + urgency + switching costs), potentially producing the first mega-outcomes in robotics.[3]
6) A useful “stack” view of where things are going
- Hardware cost compression enables scale.[2]