In our previous article, we explored how the high-level intelligence of Vision-Language Models (VLMs) translates directly into physical movement through Google DeepMind’s RT-2. We also highlighted key challenges that remain to be overcome, such as movement coarseness stemming from quantization errors and latency limitations in real-time control.

Post-RT-2, modern robotics stands at a massive inflection point. Where classical robotics once partitioned perception, planning, and control into strictly isolated modules, contemporary architectures treat physical action as a direct output generated by massive models. The boundary between semantic understanding and motor execution is steadily blurring.

Yet this transition does not mean traditional control architectures are vanishing entirely.

Instead, current developments reflect a reconfiguration rather than a complete collapse of the existing stack. Massive foundation models interpret overall context and form intent, while lower layers leverage battle-tested controllers and safety mechanisms to execute physical motor commands.

The future of Physical AI seems less about ‘a single model solving everything’ and more about ‘massive intelligence and fine-grained control collaborating through novel architectures.’

How are researchers and practitioners working to boost response speeds and achieve high-precision physical control? In this article, we examine four design approaches bridging thought and action in robotics, along with the real-world engineering trade-offs encountered along the way.

---

Post-RT-2: Four Architectures Diverging in Physical AI

Modern robot AI architectures split broadly into four distinct design paths based on how they bridge semantic reasoning with physical control.

① Isolated Planner Architecture: ‘Separating Head from Hands’

This is the most traditional yet widely adopted, battle-tested approach in industrial settings, strictly partitioning reasoning (planning) from execution (control). Represented by models such as SayCan, a massive model formulates only the macro-level plan specifying what needs to be done. Fine-grained motor execution is offloaded entirely to lower-level, pre-trained skill libraries.

• Strengths: High system stability and clear step-by-step interpretability, making it easy to isolate and debug specific failing modules.
  
  
• Limitations: Rigidity. The robot remains smart only within pre-programmed skill boundaries, struggling when encountering unfamiliar environments.

② Discrete VLA Architecture: ‘Learning Actions as Vocabulary’

Pioneered by RT-2 and advanced by the open-source community through OpenVLA, this approach discretizes continuous motor actions into tokens and embeds them directly into a large model’s vocabulary dictionary.

OpenVLA brings closed proprietary architectures like Google’s RT-2 into the open-source ecosystem, pairing a Llama-based language backbone with a vision encoder and fine-tuning on diverse robotic datasets. By leveraging fine-tuning and compression techniques, it enables multi-vendor hardware to tap into web-scale common sense.

• Strengths: Web-scale common sense flows directly into physical trajectories, offering high generalization such as selecting a metal cup as a substitute hammer without explicit prior programming.
  
  
• Limitations: Quantization errors inherently arise when continuous physical reality is compressed into discrete tokens, limiting sub-millimeter force control and high-precision assembly.

③ Continuous Trajectory Generation Architecture: ‘Physics as Continuous Flow’

Rather than tokenizing actions, this paradigm directly generates smooth, continuous motor trajectories. Diffusion Policy laid the groundwork for this approach, with Physical Intelligence’s π0 (pi-zero) recently emerging as a prominent representative.

This family excels at handling multimodal distributions inherent in real-world environments. While token-based methods risk losing information when compressing multiple paths into a single sequence, probabilistic formulations naturally capture diverse candidate trajectories. Models like π0 utilize a VLM for high-level reasoning while generating actions via continuous flow, demonstrating strong potential for precision control.

• Strengths: Enables precise, high-frequency control aligned with physical laws, seamlessly performing complex, non-rigid manipulations such as pouring liquids or folding cloth.
  
  
• Limitations: Iterative sampling over probability distributions creates significant computational overhead, making system architecture complex and compute-intensive.

④ Dual-System Architecture: ‘Slow Brain Meets Fast Spine’

Emerging as a leading standard in modern humanoid robotics, this approach mirrors human cognition by coupling System 2 (slow, deliberate reasoning) with System 1 (fast, reflexive reaction), combining two neural networks operating at different frequencies into a single embodied system.

A massive upper-level VLM interprets broad context and compresses task intent down to a lightweight, high-speed motor policy, which executes real-time motor commands at ultra-high frequencies. Architectures like Figure AI’s Helix and NVIDIA’s GR00T N1 leverage this design to precisely control high-degree-of-freedom humanoid bodies, dividing labor efficiently between networks with distinct computational speeds.

• Strengths: Simultaneously achieves broad contextual reasoning and real-time motor agility, scaling exceptionally well to complex, high-DoF robotic bodies.
  
• Limitations: High engineering complexity in synchronizing two systems operating at different execution frequencies and co-training them organically.

---

Four Comparative Axes and the Physical AI Trilemma

Regardless of their apparent complexity, all modern robotic AI paradigms aim to solve one core engineering challenge: how to translate intelligence built on web-scale data into physical robotic execution. To address this problem, engineers evaluate four key design dimensions to strike an optimal trade-off.

Axis 1. Action Representation: Discrete Tokens vs. Continuous Trajectories

This dimension reflects a fundamental architectural choice between modeling movement as sequences of discrete words or as smooth continuous trajectories.

• Discrete Token Approach (RT-2, OpenVLA): Discretizes robot motion into text tokens akin to forming sentences. This allows models to directly leverage massive web-scale text and knowledge, offering exceptional scalability for understanding novel instructions.

• Continuous Trajectory Approach (π0, Diffusion Policy): Recognizes physical motion as an inherently uninterrupted flow. By generating continuous trajectory curves, it naturally captures multimodal action distributions, yielding smoother and more precise motion.
  

Axis 2. Model Responsibility Scope: End-to-End Neural Control vs. Hybrid Safety Guards

This axis defines how much end-to-end control is delegated to neural networks, spanning from high-level perception down to low-level motor actuation.

• Fully end-to-end architectures eliminate information loss between reasoning and execution. However, they present black-box interpretability challenges and heighten safety risks during unexpected anomalies.

• To mitigate this, hybrid designs integrate classic physics-grounded safety filters beneath massive AI models. While the neural backbone retains flexible reasoning, deterministic low-level controllers safely catch critical execution errors at the end-effector.
  

Axis 3. Data Strategy: Standardization vs. Raw Ingestion

A robot model’s capability ultimately hinges on data curation and training strategies, highlighting a fascinating architectural divergence between OpenVLA and π0.

• OpenVLA’s Strategy: Aggregates heterogeneous robot datasets globally into a foundational baseline. To accommodate diverse embodiments and joint configurations, data is preprocessed and standardized into unified tokenization formats.

• π0’s Strategy: Ingests raw data directly without spatial or temporal standardization across differing robot embodiments. Utilizing advanced training techniques like Flow Matching, it preserves embodiment-specific physical characteristics and native motor control frequencies without distortion.

Axis 4. The Physical AI Trilemma: Practical Deployability

When deploying robots into real-world applications, engineers encounter a classic trilemma. Satisfying all three fundamental capabilities within a single monolithic architecture remains extraordinarily difficult:

• Generality: Adaptability to unseen objects, environments, and novel user instructions (the primary strength of massive foundation models).
• Dexterity: Nuanced physical force regulation and precise manipulation aligned with physical laws (the primary strength of trajectory generation models).
• Deployability: Low computational overhead and high-frequency real-time execution speeds in physical environments (the primary strength of traditional control stacks).

Ultimately, no single architecture holds a universal monopoly on correctness. The true engineering question becomes: ‘Which objective takes priority given our specific deployment constraints?’ High-precision manufacturing environments demand low latency and sub-millimeter dexterity, whereas domestic service robots prioritize open-ended generality to handle novel household objects.

---

Four Persistent Bottlenecks in Physical AI

Even with an immaculate architectural blueprint, every model eventually collides with the harsh realities of the physical world. These challenges do not stem merely from flaws in specific model designs, but rather from the fundamental constraints of physics. Here are the four primary bottlenecks facing the field today.

Bottleneck 1. The Resolution of Intelligence: Discrete Symbols vs. Continuous Physics

Representing robotic actions as text tokens enables flexible, dialog-like expansion, yet it sacrifices physical precision. Conversely, generating continuous action trajectories recovers fine-grained dexterity, but causes spatial computational complexity to scale exponentially. The oldest debate in artificial intelligence, whether intelligence should be expressed through discrete symbols or continuous physical flows, continues to collide within modern robotics.

Bottleneck 2. Data Asymmetry: Finding the Golden Ratio of Pre-training and Post-training

Web-scale internet datasets are vast, yet they lack tactile embodiment. Conversely, physical trajectory datasets contain raw physical ground truth, but remain scarce and expensive to collect. To bridge this gap, researchers rely on a two-stage paradigm: Pre-training followed by Post-training alignment. Models first absorb broad common sense from web data, then align low-level torque and force dynamics via simulators or robot-specific fine-tuning. The exact mixing ratio of web and physical datasets remains a tightly guarded trade secret among leading Physical AI teams.

Bottleneck 3. The Duality of Generalization: Smart Reasoning Meets Clumsy Actuation

Inside a robot, high-level semantic generalization (understanding novel objects) frequently desynchronizes with low-level motor control generalization (adapting to unfamiliar friction, compliance, or mass). Recent models like π0.5 demonstrate impressive multi-task co-training, successfully navigating completely unseen home environments to manipulate household objects. Yet achieving high first-pass success rates consistently remains a hurdle, proving that pairing smart cognitive reasoning with robust physical control is still an unsolved engineering challenge.

Bottleneck 4. Latency and Responsiveness: Autoregressive Tokenization vs. Action Chunking

In real-world physical environments, a 0.1 second inference delay is not merely a software lag, but a potential physical crash. Autoregressively generating action tokens one by one imposes severe latency constraints on high-frequency closed-loop execution. To bypass this bottleneck, diffusion-based models adopt Action Chunking as a practical survival mechanism, predicting and transmitting a continuous sequence of actions spanning the next 1 to 2 seconds in a single forward pass.

---

Where Is Physical AI Heading Next?

Will future robotics converge into a single monolithic foundation model that entirely dissolves the boundaries between perception, planning, and control?

Recent breakthroughs like Google DeepMind’s Gemini Robotics, which pairs Gemini 2.0 with physical motor execution, highlight this possibility. Performing fine-grained tasks such as origami and card handling seamlessly across novel hardware platforms proves the immense generalization power of a unified cross-modal representation. It almost feels as if the traditional robotic stack is on the brink of total collapse.

Yet moving from lab benchmarks to real-world industrial deployments tells a different story. Operating physical robots requires far more than semantic comprehension. It demands millisecond-level reaction times, clear root-cause interpretability, and uncompromising physical safety.

Consequently, rather than consolidating into a single universal paradigm, physical AI is likely to branch into four distinct evolutionary vectors aligned with specific deployment constraints:

Vector 1. Generalist Foundation Models (Generalist VLA)

Leveraging massive foundation models through prompt engineering or lightweight fine-tuning to achieve zero-shot or few-shot adaptation across heterogeneous hardware. This represents a direct attempt to bring the scaling laws of foundation models to the robotics market.

Vector 2. Layered Hybrid Architectures

High-level semantic intent like ‘clear the hazardous object’ is processed by a massive VLM, while real-time force regulation and safety fallback mechanisms are handled by high-frequency local controllers. In industrial settings where deterministic execution and safety verification are mandatory, this hybrid structure represents the most pragmatic standard. On-device lightweight models like SmolVLA follow this philosophy.

Vector 3. Efficient Action Representations

Research like π0-FAST, which compresses continuous trajectories into frequency domain representations, along with Latent Action Space exploration, targets this space. This is a technical race to engineer an optimal action grammar capable of preserving both web-scale semantics and fine-grained continuous control without information loss.

Vector 4. Optimization of Data Recipes

The bottleneck in Physical AI is shifting rapidly from model architecture to data curation. The core challenge lies in mixing and aligning web-scale vision-language datasets with embodied physical trajectories. Developing clever synthesis pipelines that blend photorealistic simulator data with real-world physical interactions will serve as a definitive competitive edge.

Robotics must absorb the massive scaling trends developed in pure software AI while simultaneously satisfying far stricter physical constraints and safety requirements. If a language model emits a hallucination, it can self-correct in subsequent sentence generation. In physical robotics, however, a single erroneous motor command can cause severe hardware failure or physical injury.

The reason we refer to this control boundary as ‘the gap between thought and action’ is not merely due to technical limits. It is because the complex constraints bridging semantic reasoning to motor actuation, including perception, planning, control, safety, and data availability, are all compressed into this single inflection point.

Yet this gap is no longer an insurmountable barrier. From planning to policy, and from tokens to continuous trajectories, the connective architectures linking mind to movement are being reimagined in fascinating ways.

Ultimately, true intelligence in physical AI will not be measured merely by how much common sense a model comprehends, but by how reliably and safely it converts that understanding into physical execution at the fingertips.