AI, Robotics, and Embodied Systems
Foundation models trained on internet-scale data are transferring into robots — closing the gap between digital and physical AI, and reopening the long-deferred dream of general-purpose machines.
Key facts
- Open X-Embodiment combines data from 22 robot embodiments across 21 institutions.
- Humanoid funding crossed $10B cumulative by 2025.
- Sim-to-real with domain randomization is now standard for manipulation training.
- Waymo crossed 100,000+ paid driverless trips per week in 2025.
- Figure, 1X, and Agility have all signed paid commercial pilots.
Robot Foundation Models
RT-2, OpenVLA, Octo, π0 (Physical Intelligence), and Helix (Figure) demonstrate that vision-language-action (VLA) models can serve as general robot policies, generalizing across tasks and embodiments with relatively little task-specific data.
Pretraining on internet-scale image-text data, then fine-tuning on robotics demonstrations, has become the dominant paradigm — analogous to how LLMs are post-trained.
Humanoids
Figure (02), 1X (Neo), Tesla Optimus, Agility (Digit), Apptronik (Apollo), Unitree (G1, H1), and Chinese players (Fourier, XPeng IRON, Unitree) have moved from prototype videos to early commercial pilots in warehouses and labs.
The economic case rests on general-purpose physical labor priced competitively against human shift-work, plus the leverage of bipedal form-factor in existing human environments.
The Data Problem
Robotics lacks an internet-scale data source. Sim-to-real transfer (using simulators like NVIDIA Isaac), teleoperation farms, video pretraining (Ego4D, Open-X), and shared cross-embodiment datasets (Open X-Embodiment, DROID) are bridging the gap.
Synthetic data plus real-world fine-tuning is becoming standard; pure end-to-end real-world training remains expensive and slow.
Safety, Standards, and Workplaces
ISO 10218 (industrial robots), ISO 15066 (collaborative robots), and new humanoid-specific standards (ISO/IEC working groups, ANSI/RIA work) define safe interaction with humans.
Insurance, OSHA-style workplace rules, and liability frameworks are not yet mature for autonomous humanoids outside structured environments.
Realistic Outlook
Structured-environment deployments (warehouses, factories, controlled retail) are scaling now. General-purpose home humanoids remain years away — dexterity, long-horizon reliability, and cost are the binding constraints.
Autonomous vehicles (Waymo, Zoox, Wayve) are the most mature embodied-AI category by hours of deployment.
Frequently asked
Are humanoid robots really coming?
+
Limited commercial pilots exist today in warehouses and labs. General-purpose home humanoids remain years away due to dexterity, reliability, safety, and cost constraints.
What's the bottleneck — software or hardware?
+
Both. Hardware costs are falling fast; software for long-horizon dexterity and safe human interaction remains the harder problem.
Will robots take blue-collar jobs?
+
Selectively, in structured environments. Unstructured trades (plumbing, framing, in-home care) remain very hard for embodied AI.
Sources & further reading
Continue in this series
Healthcare
AI in Medicine and Diagnostics
Research Acceleration
AI for Scientific Discovery
Tutoring
AI in Education and Personalized Learning
Co-creation
AI and Human Creativity
Enterprise
AI in Knowledge Work and Productivity