1. LLM, AGI, and Physical AI
The AGI race, starting with LLMs, is a journey to achieve human-level or beyond intelligence. As of 2026, it seems achievable in the near future.
But this intelligence is limited to Cognitive Intelligence.
- Coding, math, reasoning, research, literature…
The journey to solve Physical Intelligence is a different dimension of problem.
Just as ChatGPT changed the world, there’s an expectation that Physical AI will transform the world of physical labor.
2. Latest Demos (2026.01)
CES 2026
Boston Dynamics Atlas + LBM
The star of this CES, Boston Dynamics' ATLASDoes this demo have “intelligence”?
Boston Dynamics, the pioneer of Classical Robotics, is also transitioning to Physical AI.
- 450M Diffusion Transformer: Co-developed with Toyota Research Institute (TRI)
- Whole-body single model control: Walking + manipulation integrated
- Deformable object manipulation like rope tying, cloth unfolding
→ Details: LBM (Large Behavior Model)
Sharpa CraftNet
Sharpa unveiled CraftNet and the North humanoid.
- VTLA (Vision-Tactile-Language-Action): First commercial model integrating tactile sensing into VLA
- Pinwheel folding, playing card handling, and other demos requiring tactile sensing
- CES 2026 Innovation Award winner
Figure Helix 02
Figure AI’s Helix 02 is the first fully autonomous whole-body humanoid.
- 61 consecutive actions over 4 minutes: No resets, no human intervention
- System 0/1/2 architecture: High-speed control up to 1kHz
- Replaced 109,504 lines of C++ code with a 10M parameter neural network
3. What is Physical AI?
An AI system based on end-to-end VLA models that performs generalist physical tasks impossible with past rule-based approaches
Why is Physical AI Getting Attention?
What was impossible before is now possible.
| Past (Classical Robotics) | Present (Physical AI) |
|---|---|
| Only predetermined actions | Adapts to various situations |
| Couldn’t fold laundry | Can fold laundry |
| Struggled with deformable objects | Can handle plastic wrap, ropes |
Classical Robotics vs Physical AI
| Aspect | Classical Robotics | Physical AI (VLA) |
|---|---|---|
| Architecture | Modular (perception → planning → control) | End-to-end integrated |
| Learning | Rule-based + partial ML | Data-driven full learning |
| Generalization | Bound to training environment | Zero-shot generalization |
| Knowledge | Domain-specific | Inherits World Knowledge (LLM/VLM) |
→ Details: Definition of Physical AI
4. VLA, RFM, LBM Terminology
Evolution from LLM to VLA
LLM → VLM → VLA
Language → + Vision → + Action| Term | Full Name | Description |
|---|---|---|
| VLA | Vision-Language-Action | Model integrating vision + language + action |
| LBM | Large Behavior Model | Action expressed as Behavior. Same as VLA |
| RFM | Robot Foundation Model | Foundation Model for robots |
Why VLA is Special
LLMs have common sense. So VLAs have common sense too.
- Inherits World Knowledge from LLMs that learned all internet knowledge
- Can work at a different cafe, with new menu items
- Can handle packages of various shapes, clothes of various designs
→ Details: What are RFM & VLA?
5. Data Scaling Problem
There are reasons why VLA can’t simply follow LLM’s success formula.
Difference Between LLM and VLA
| Aspect | LLM | VLA |
|---|---|---|
| Data Source | Internet (infinite) | Real robot actions (limited) |
| Collection Cost | Low | High |
| Evaluation | Can be automated | Requires physical robot operation |
Solution Attempts
| Approach | By | Description |
|---|---|---|
| Teleoperation | Tesla, Google, PI | Direct data collection |
| Simulation | NVIDIA | Omniverse + Cosmos |
| Community | HuggingFace | Open-source collaboration |
| World Model | 1X, NVIDIA | Synthetic data generation |
→ Details: Action Data Scaling Problem
6. VLA & RFM Progress
Convergent Evolution: System 1/2 Architecture
In 2025, different research groups independently arrived at similar structures.
| System | Role | Frequency |
|---|---|---|
| System 2 | High-level planning, language/vision understanding | 7-10 Hz |
| System 1 | Low-level motor control | 100-200 Hz |
Models that adopted this:
- GR00T N1.6 (NVIDIA)
- Figure Helix (Figure AI)
- Gemini Robotics (Google DeepMind)
Convergent Evolution: Continuous Action Generation
RT-2’s “Action as Language” (discrete tokens) → Flow Matching / Diffusion (continuous values)
| Model | Method | Features |
|---|---|---|
| π0 | Flow Matching | 50Hz control |
| GR00T N1 | Diffusion Transformer | Dual system |
| SmolVLA | Flow Matching | 450M lightweight |
→ Details: VLA & RFM Progress
7. Physical vs Cognitive Intelligence
Moravec’s Paradox
“High-level reasoning requires relatively little computation, but low-level sensorimotor skills require enormous computational resources.”
| Easy (for AI) | Hard (for AI) |
|---|---|
| Beat chess world champion | Take keys out of pocket |
| Beat Go masters | Wash dishes |
| Write complex equations | Pick up fruit |
AI that beats chess champions was created in 1997, but taking keys out of a pocket still doesn’t work well.
Why is Physical Intelligence Hard?
- Evolutionary perspective: Motor/sensory/perception evolved over ~1 billion years, abstract thinking over ~millions of years
- Dimensionality problem: Text is abstracted low-dimension, physical world is high-dimension + real-time interaction
- Learning difference: Humans learn through experience, VLAs mainly through imitation learning
→ Details: Physical vs Cognitive Intelligence
8. Conclusion: It Still Seems Possible
Reasons for optimism:
- Much of human labor doesn’t necessarily require tactile sensing
- Robots can add senses humans don’t have (depth cameras, etc.)
- Robots can make movements impossible for human bodies
- As Tesla FSD showed in autonomous driving, implementation can differ from human approach
Humanity, having learned so much from LLMs, seems likely to find the answer soon.
Learn More
Fundamentals Guide
Insights
Key Models
- π0 / π0.5 - Physical Intelligence
- GR00T N1 - NVIDIA
- Figure Helix - Figure AI
- LBM - Boston Dynamics + TRI
- CraftNet - Sharpa