Over the past six months, the pace of the embodied AI field has clearly accelerated.
Physical Intelligence released π₀.7, Google DeepMind launched Gemini Robotics 1.5 and Robotics-ER 1.5, NVIDIA’s Physical AI full stack moved from concept to open source, and Figure AI’s humanoid robot completed over 11 months of deployment at the BMW factory. On the Chinese side, Unitree completed its IPO in August 2026 (valued at approximately $9 billion), while Agibot and Galaxy General continued to expand their financing rounds, with humanoid robots beginning to move from lab prototypes toward small-batch production.
This article attempts a foundational survey: what technology stack is embodied AI forming? What problems are the major players solving? And what stage have they reached? “From VLA to world models” is not a linear technology evolution path, but rather a main thread for observing the embodied AI technology stack. No predictions, no rankings — just drawing the technology map clearly.
Three Core Questions: Policy, Dynamics, Infrastructure
Rather than dividing embodied AI participants into “three schools,” it is more useful to ask three different questions: how to make a robot act? how to predict the future? how to obtain training data and environments? These three questions correspond to three different technical layers, and a complete embodied AI system typically needs to address all of them.
Policy: How to Make a Robot Learn to Act?
Core idea: Map from perception (vision, language, proprioception) directly to action, or through intermediate planning to action output.
Representatives: Physical Intelligence (π₀ series), Google DeepMind (RT-2 → Gemini Robotics 1.5), NVIDIA GR00T
Physical Intelligence’s π₀ series is currently the most representative work in the VLA approach. From π₀’s flow matching continuous action generation, to π₀.5’s discrete-continuous hybrid recipe, to π₀.7’s context-rich steering + visual sub-goals — this technical line has been continuously iterating on action interface design.
π₀.7 is particularly noteworthy because it demonstrates that VLA has not simply rejected world models — it is gradually absorbing predictive components. π₀.7’s visual subgoal is generated by a lightweight world model — meaning that π₀.7’s core policy interface is still not an explicit action-conditioned rollout interface, but the system has already introduced structured prediction of future visual states through an external predictive component. The cross-embodiment zero-shot T-shirt folding demonstrated by π₀.7 is a landmark result, but it should be noted that this is still a result under a specific evaluation protocol, and there is considerable distance to general household manipulation.
Google DeepMind’s approach has evolved from RT-2’s vision-language model adapted for robotics, through Gemini Robotics, and further expanded to Robotics-ER 1.5 and subsequent versions. The key change is not simply “a bigger VLM,” but rather gradually forming a layered system composed of VLA, embodied reasoning, and high-level task planning. Gemini Robotics 1.5 is no longer simply “VLM → robot policy” — the official description positions Robotics-ER 1.5 for high-level embodied reasoning / planning, while Robotics 1.5 handles vision-to-motor-commands mapping. The direction of this evolution is: letting semantic understanding, physical reasoning, and action generation each play their role in a layered architecture.
NVIDIA’s GR00T N1/N2 also belongs at this layer — it is an open-source humanoid robot foundation model providing policy-level capability.
The advantage of this approach is strong semantic generalization — the semantic knowledge VLA inherits from internet pre-training enables it to handle novel objects and novel instructions. The limitation is: the core interface of traditional VLA policy remains observation / instruction → action, rather than explicitly outputting action-conditioned future states available for rollout — but as π₀.7 and Gemini Robotics 1.5 demonstrate, recent systems are gradually adding predictive, subgoal, and high-level planning modules.
Dynamics: How to Predict the Future?
Core idea: Learn action-conditioned dynamics in latent space, then use the prediction model for control through imagination, MPC, or value estimation. The “Dynamics” here refers to model components responsible for predicting how future states change, not a narrow model architecture; a complete world model may also include interfaces for reward, termination, observation prediction, and more.
Typical approaches: DreamerV3 (RSSM + imagination), TD-MPC2 (latent dynamics + MPC); extending to foundation-model scale includes NVIDIA Cosmos and other world foundation models.
DreamerV3 learns a dynamics model in latent space through RSSM, then rolls out trajectories in imagination to train an actor-critic policy. This approach has a clear advantage in sample efficiency — complex skills can be learned from a small amount of real interaction data.
TD-MPC2 takes a more concise approach: encoder + MLP ensemble dynamics + latent-space MPC + Q-function ensemble. Without the dual-track structure of RSSM, it demonstrates potential in cross-task scaling — evaluated on 104 continuous control tasks across 4 domains, and further demonstrating that a single 317M-parameter agent can be trained across 80 tasks (for detailed architecture analysis of TD-MPC2, see a previous article).
The advantage of this approach is sample efficiency, and the explicit use of action-conditioned dynamics for prediction, planning, or value estimation. It is important to note that latent predictive model ≠ accurate physical simulator — the advantage of latent prediction models is not “high-precision physical prediction,” but rather providing a predictive interface usable for planning and value bootstrap. The limitation is: world model prediction quality degrades with rollout length, and reliability for long-horizon tasks remains a bottleneck. For traditional world model approaches centered on latent dynamics + model-based control (such as Dreamer, TD-MPC2), language grounding and open-world semantic generalization are typically not their primary optimization objectives, and in these capabilities they generally do not match VLAs built on foundation models.
NVIDIA Cosmos also belongs at this Dynamics layer, but a distinction is needed: Dreamer / TD-MPC2 are closer to control-oriented latent dynamics, while Cosmos follows a foundation-model approach oriented toward physical world modeling, generation, and policy model training — the two are not the same kind of world model. Cosmos can be used for physical world prediction, world generation, and training/generation of action / policy models. Its positioning has gone far beyond “synthetic data generator” — it now encompasses world models, post-training, data processing, and evaluation as a Physical AI development stack.
Infrastructure: How to Obtain Training Data and Environments?
Core idea: Provide simulation platforms, data toolchains, and foundation models to support training and evaluation of the Policy and Dynamics layers above.
Representatives: NVIDIA (Isaac / Omniverse), World Labs (Marble), teleoperation data collection and dataset curation systems
Infrastructure is not just simulation — it also includes robot data collection, teleoperation, dataset curation, synthetic data generation, and evaluation.
At this layer, NVIDIA is building a Physical AI full stack: Isaac / Omniverse provides simulation and data generation infrastructure, GR00T provides robot foundation models (Policy layer), and Cosmos provides world foundation models with data generation, prediction, and post-training capabilities (Dynamics layer). NVIDIA does not use mass-produced robots under its own brand as core products, but instead aims to serve the robotics ecosystem through its full-stack hardware, software, and development tools — a positioning with strong leverage effects.
Spatial World Models: World Labs
World Labs deserves separate discussion, as its technical positioning differs from the infrastructure players above.
World Labs’s Marble is a multimodal world model that generates explorable, editable 3D worlds from inputs such as text / image / video. It currently sits closer to spatial intelligence / generative world modeling rather than a complete robot policy stack.
World Labs Marble
Generate / reconstruct 3D world
↓
Becomes simulation / spatial environment asset
↓
Robotics / simulation downstream
Marble has the potential to become an important environment generation component in the sim-to-real chain, helping improve the diversity and realism of simulation environments. But it is not itself a robot policy system — rather, it provides spatial intelligence infrastructure for downstream Policy and Dynamics layers.
Humanoid Robots: The Bet on Hardware Form Factor
One of the most prominent trends in 2026 is the collective sprint in humanoid robots.
Figure AI’s Figure 02 underwent 11 months of deployment at the BMW Spartanburg factory, accumulating over 1,250 hours of runtime and contributing to the production of 30,000+ X3 vehicles (all figures from Figure’s official disclosure). Figure’s technical approach combines end-to-end learning (in partnership with OpenAI) and traditional control. Among publicly disclosed humanoid robot cases, this is one of the closer examples to long-term real deployment.
Tesla Optimus continues to iterate on hardware design, targeting internal factory deployment. Tesla’s advantage lies in its vertical integration capability — its own factories provide testing environments, its own chips (Dojo/FSD) provide training compute, and its own AI team provides algorithms.
1X Technologies’ NEO series targets general-purpose service scenarios, with a technical approach leaning toward end-to-end learning.
Chinese companies are investing particularly intensively in this direction. Unitree pivoted from quadruped robots to humanoid, with fast hardware iteration and strong cost control, and completed its IPO in August 2026 (pricing corresponding to approximately $9 billion valuation). Agibot and Galaxy General have continued to expand their financing — Galaxy General's valuation has reportedly reached the $3 billion level.
Humanoid robots are a high-risk, high-reward bet. The advantage is: humanoids can adapt to environments designed for humans (stairs, door handles, tools). The risk is: the engineering complexity of humanoids is far greater than that of specialized-form-factor robots, and a significant portion of publicly demonstrated humanoid robot demos still rely on teleoperation, predefined routines, or controlled policies, with considerable distance remaining to autonomously completing complex tasks in open environments.
Chinese Embodied AI: Fast Catching Up and Differentiation
The Chinese embodied AI track has shown several characteristics over the past six months.
Accelerating financing and IPOs. Unitree completed its IPO, and multiple leading companies have entered unicorn territory. Capital is shifting from “investing in concepts” to “investing in deployment.”
Outstanding hardware capability. China’s supply chain advantage in robot hardware (motors, reducers, sensors) is translating into whole-machine advantage. Unitree’s cost control capability is competitive on a global scale.
Gap remains in software/algorithms. Based on publicly available papers, models, and benchmarks, Chinese embodied AI companies still have a gap compared to leading research teams like Physical Intelligence and Google DeepMind in areas such as general VLA foundation models, cross-embodiment data scale, and public technical influence; however, the true size of this gap is difficult to accurately quantify based solely on public materials. At the same time, the open-source ecosystem, public papers, and talent mobility are lowering some technical barriers, but whether these factors will ultimately translate into equivalent competitiveness in foundation model capabilities still requires more public benchmark validation.
Application scenario differentiation. Some Chinese companies place greater emphasis on clearly defined scenarios such as industrial, warehouse, and commercial services deployment, while American leading companies simultaneously bet on general-purpose humanoids, foundation models, and cross-scenario generalization. Both strategies involve tradeoffs — scenario focus enables faster commercial validation, but accumulation of generalizability may be slower.
Several Technical Trends Worth Watching
Setting aside specific companies, there are several technical trends that deserve continued attention.
Trend One: VLA and World Models Are Forming Hybrid Architectures
The relationship between VLA and world models is no longer simply “converging from both ends.” More accurately, the industry is gradually forming one possible layered structure (note: this is not an industry-established canonical architecture — actual system module combinations may differ significantly):
Semantic / Reasoning
↓
VLA
↓
Predictive / World Model
↓
Planning / Value
↓
High-frequency Control
↓
Embodiment
↓
Real World
In this layering, VLA provides semantic prior / task understanding, world models provide predictive interfaces, and low-level controllers / policies provide high-frequency action execution. For example: π₀.7’s visual subgoals come from a lightweight world model; Gemini Robotics 1.5 separates embodied reasoning from VLA into distinct layers; NVIDIA Cosmos provides world foundation models + synthetic data; GR00T provides foundation policy; TD-MPC2 provides latent planning.
This suggests the core competitive question may not be “VLA vs World Model,” but rather: who handles semantic abstraction, who handles prediction, who handles control. As of now, no single system simultaneously possesses mature language grounding, action-conditioned prediction, and high-frequency continuous control — but multiple systems are approaching this goal by composing different modules.
Trend Two: Simulation Becomes a Key Component of Mainstream R&D
Nearly all major players are using simulation at scale for training or data augmentation. The use of simulation platforms like NVIDIA Isaac Sim, MuJoCo, and Isaac Lab has become an important component of mainstream robot R&D workflows. The sim-to-real gap still exists, particularly in fine manipulation, contact dynamics, and long-tail environments.
Trend Three: Data Is Becoming a Key Differentiator
An increasingly obvious trend is that pure model architecture differences are becoming less likely to form decisive advantages, while the importance of data scale, data diversity, and training recipes is rising. This is not to say architecture doesn’t matter — rather, the performance bottleneck in the foundation model era is increasingly shifting toward data / compute / embodiment coverage. Whoever can obtain more diverse, higher-quality robot interaction data covering broader embodiment and task distributions is more likely to form an advantage. This explains why teleoperation data collection, synthetic data generation, data quality filtering, and other “data engineering” directions are receiving more attention.
Trend Four: The Gap from Demo to Deployment
Most publicly demonstrated robot capabilities remain “demo-level” — completing specific tasks in controlled environments. There is a massive engineering gap between demo and reliable deployment — “completing a task” and “running continuously for 8 hours” are not the same problem at all. This gap spans multiple levels: task success → robustness → failure recovery → long-horizon autonomy → fleet-level reliability → maintenance → economic viability. Figure AI’s 11-month deployment test at BMW is among the closer examples to long-term real deployment in publicly available information.
What Does This Map Mean?
If we summarize the current technology map of embodied AI:
Embodied AI
│
┌─────────────────────┼─────────────────────┐
│ │ │
Policy Dynamics Infrastructure
│ │ │
π₀ / Gemini RSSM / TD-MPC2 Isaac / Cosmos
GR00T / ... Video WMs World Labs
│ │ │
└──────────────┬──────┴──────────────┬──────┘
│ │
Data / Simulation / Evaluation Embodiment
│ │
└──────────┬──────────┘
│
Robot Systems
│
┌────────┴────────┐
│ │
Industrial General
deployment humanoid
What is truly competing is not three isolated approaches, but how policy, dynamics, data, simulation, and embodiment ultimately combine.
Several assessments:
First, technical approaches have not yet converged. The Policy, Dynamics, and Infrastructure layers each have unsolved problems. The final system will likely be a hybrid architecture — VLA providing semantics, world models providing prediction, low-level control providing execution — but a clear convergence direction is not yet visible.
Second, software, data, and models are becoming increasingly important sources of differentiation. As humanoid robot hardware gradually sees more standardized components and a maturing supply chain, differentiation sources are increasingly extending to software, data, and models. But this transition is still early — actuator, hand, force sensing, whole-body control, and other hardware capabilities still significantly affect actual robot performance, and hardware + software co-design may in fact become increasingly important.
Third, the gap from demo to deployment is the biggest current challenge. Most public results remain “it works under specific conditions” rather than “it runs reliably in real environments.” The key to solving this problem may not be larger models, but better data, more robust policies, a more mature sim-to-real pipeline, and systems engineering capabilities in failure recovery, monitoring, and fleet-level reliability.
Fourth, China has clear advantages in hardware supply chain, manufacturing, and cost control, and is also very active in commercialization exploration. Unitree’s IPO and the financing progress of multiple companies demonstrate that hardware and commercialization capabilities have been validated by the market. However, in areas such as large-scale reliable deployment and general-purpose foundation models, publicly available evidence is not yet sufficient to demonstrate equivalent advantages. The open-source ecosystem, public papers, and talent mobility are lowering some technical barriers, but the true size of the gap still requires more benchmark data to quantify.
This is a snapshot at a point in time. The embodied AI field is changing rapidly; this map may need to be redrawn in six months.
Next up is the follow-up to the VLA series — an analysis of the relationship between VLA and world models, open questions, and three assessments.
Comments