The year 2026 is shaping up to be the moment Physical AI finally steps out of the lab and into factories, warehouses, and even sidewalks. Vision-Language-Action (VLA) models are the brains, humanoid robots are the brawn, and a fierce geopolitical race is unfolding between the United States and China. But where does Russia fit into this picture? A recent deep dive on Habr provides one of the most comprehensive breakdowns of the global Physical AI landscape today, and the stakes are higher than ever.
The VLA Revolution: Brains That See, Think, and Act
Traditional robotics relied on rigid programming: a robot arm follows a predefined path, a drone flies a GPS route. Physical AI flips that script. VLA models — short for Vision-Language-Action — are large neural networks trained on massive datasets of text, images, and robot trajectories. They allow a robot to perceive its environment, understand a natural-language command like “pick up the red cup,” and execute the action without explicit step-by-step code.
Google DeepMind’s RT-2 (Robotic Transformer 2) was an early milestone, demonstrating that a model trained on web data could generalize to new robotic tasks. In 2026, the ecosystem has exploded. Companies like Covariant, Physical Intelligence, and startup Skild AI are deploying VLA models in commercial settings. What’s new this year is the shift from simple pick-and-place to multi-step manipulation: folding laundry, assembling electronics, even cleaning a kitchen.
As the Habr article notes, the key bottleneck is data. Training a VLA model requires millions of real-world robot interactions, which are expensive and slow to collect. Simulation-to-reality (sim2real) transfer has improved, but physical data remains the gold standard. Several Chinese labs have circumvented this by using human teleoperation to generate action trajectories en masse — a approach the article calls “data farming.”
Humanoids: Finally Leaving the Lab?
Humanoid robots have been a sci-fi staple for decades, but 2026 might be the year they become a commercial reality. The Habr analysis points to several tipping points:
- Tesla Optimus Gen 3 has started limited deployment in Tesla factories for material handling. Elon Musk has promised broader availability by 2027, but reliability and cost ($20,000–$30,000 per unit) remain hurdles.
- Figure AI secured a massive partnership with an unnamed automaker and claims its Figure 02 can perform 80% of automotive assembly tasks autonomously.
- Boston Dynamics surprised the industry by releasing a commercial version of Atlas, now fully electric and optimized for logistics.
- Chinese players like Unitree, Fourier Intelligence, and Xiaomi are flooding the market with cheaper humanoids. Unitree’s H1 is priced under $90,000 and can carry 30 kg payloads – aggressive pricing that pressures Western competitors.
The article emphasizes a crucial trend: humanoids are being designed not as general-purpose butlers, but as task-specific workers. The software stack (VLA models + motion planning) is becoming more important than the hardware. Companies that own the best models and the largest data pipelines will likely dominate.
The US-China Race: Supply Chains, Policies, and Data Sovereignty
The geopolitical dimension of Physical AI is impossible to ignore. The Habr breakdown highlights how the US and China are pursuing fundamentally different strategies:
| Aspect | United States | China |
|---|---|---|
| Core approach | Open research (academia + big tech) | Centralized state-led push with massive subsidies |
| Data advantage | Access to diverse real-world scenarios in logistics, healthcare, etc. | Scale: thousands of human-teleoperation data stations, huge manufacturing floors |
| Hardware | Leading in high-end actuators, sensors (e.g., Tesla, Boston Dynamics) | Dominant in supply chain for motors, batteries, and cost-efficient production |
| Key players | Google DeepMind, Tesla, Figure AI, Boston Dynamics, Apple (rumored) | Unitree, Fourier, Dreame Technology, Xiaomi, Huawei Robotics (emerging) |
| Policy | Export controls on advanced chips (NVIDIA H100/B100 restrictions); CHIPS Act funding | ‘Made in China 2025’ plus local robotics hubs; less constrained by chip bans for inference |
The race is not just about robots – it’s about the foundational AI models. US companies have an edge in VLA research, but Chinese firms compensate with vast amounts of physical-world data from their manufacturing ecosystem. The article notes that several US startups have quietly set up data-collection centers in Southeast Asia to sidestep Chinese government restrictions.
What Does Russia Have?
This is where the picture gets murky. Russia’s Physical AI landscape is modest but not absent. The Habr analysis points to a few notable efforts:
- Promobot (Perm) remains the most visible Russian robotics company, producing service robots for retail and banking. Their platform is being updated with VLA-like capabilities via partnerships with academic labs at Skolkovo and MIPT.
- The Humanoid Robotics Laboratory at the Moscow Institute of Physics and Technology has developed a full-size humanoid prototype named “Fedor-2,” but it’s still far from production.
- Industrial automation – companies like Sber Robotics and Yandex Self-Driving Group have integrated computer vision and reinforcement learning into warehouse robots, but these are not general-purpose VLA systems.
- Challenges: Russia faces severe restrictions on importing high-end GPUs and advanced actuators due to sanctions. The brain drain of AI talent has also accelerated since 2022. As a result, most Russian efforts remain at the research-prototype stage.
However, the article highlights a silver lining: Russia has strong theoretical foundations in control theory and optimization (lingering from the Soviet era), and several teams are producing novel work in sim2real algorithms and energy-efficient locomotion. But without access to cutting-edge hardware and large-scale data pipelines, the gap with the US and China is widening.
What’s Next: The Convergence of VLA and Humanoids
The Habr piece concludes with a prediction: by 2030, the Physical AI market could surpass $100 billion, driven by logistics, manufacturing, and eventually domestic service. The key enabler will be the convergence of VLA models with humanoid form factors. When a robot can understand “fold the laundry in the basket” and physically do it across varied environments, the economic shift will be enormous.
For now, the US leads in model intelligence, China leads in hardware scale, and Russia is a distant contender with pockets of niche expertise. But the race is far from over – as VLA models improve, the hardware itself may become commoditized, leveling the playing field.
Final Words
If you want to dig into the full details, I highly recommend reading the original Habr article Source. It’s one of the most thorough overviews of Physical AI available in 2026, covering technical architectures, company profiles, and geopolitical implications. Whether you’re a robotics engineer or just a tech enthusiast, this space is moving fast – and understanding where the pieces fit can help you spot the next big opportunity.
Comments