As physical AI and world models continue to evolve, artificial intelligence is extending from perception and generation to understanding, reasoning, and decision-making in the physical world. For industrial manufacturing, enabling robots to understand physical environments, predict interaction outcomes, and transfer virtual training capabilities to real production lines has become a critical challenge for the deployment of embodied intelligence.
Around this technological trend, MIR Databank conducted thematic research and invited Academician Guo Song from HKUST, Wu Kan, President of Estun, and He Wenwu, Chairman of Freedo Technology for an in-depth interview to explore the technical pathways and industrial practices of embodied intelligence driven by world models entering factories.
Core Perspectives of the Three Parties: From World Models to Industrial Physical AI, Bridging the Last Mile Between Technology and Industry
HKUST | Academician Guo Song
“Industrial scenarios are the optimal testing ground for physical AI. Our team proposed the PAL (Physical Alignment Learning) framework, advocating for a data closed-loop of 'simulation pre-training — real-machine calibration — failure case feedback to simulation.' We follow a gradual industrial implementation route, prioritizing the solidification of the industrial foundation before supporting the development of humanoid and other embodied carriers.”
Estun President | Wu Kan
“Physical AI can enhance flexibility, efficiency, and safety on the manufacturing floor; this is its ultimate industrial value. Estun will leverage robot bodies, control systems, and real production line verification capabilities to drive the migration of simulation training results to real machines, prioritizing the resolution of flexible manufacturing needs in industrial scenarios.”
Freedo Technology Chairman & CEO | He Wenwu
“Industrial physical AI relies on real data and physically credible simulation environments. Freedo will undertake the prediction and training requirements of HKUST’s world models, interface with Estun’s real-machine process parameters, ensure consistency in dynamics and contact characteristics between simulation and real machines, and support the implementation of the 'simulation pre-training, real-machine calibration, case feedback' data closed-loop.”
The above represents the core claims of the three guests distilled by MIR Databank based on the entire interview content, facilitating readers to quickly grasp the technical positioning, industrial demands, and long-term strategic direction of this industry-academia-research consortium. The full Q&A transcript follows below.
In-Depth Interview: World Models and Physical AI Direction
Q1:How do world models drive embodied intelligence to truly enter factories?
We understand that this signing involves HKUST leading the exploration of the technical paradigm of world models and embodied intelligence, Estun providing industrial robot bodies and real production line verification scenarios, and Freedo providing industrial-grade simulation infrastructure support. We would like to ask the three experts to answer a core question: How do world models drive embodied intelligence to truly enter factories?
Let us first hand the main thread over to Academician Guo Song from HKUST. Academician Guo, please first define two key terms: What is a world model, what is physical AI, and why discuss them together?
Academician Guo Song: A world model is a machine's internal representation of the operational laws of the physical world. It plays two roles in robot learning. First, as an external simulator, it provides a virtual environment for data generation, policy evaluation, model predictive control, and reinforcement learning. Second, as an internal prediction module within robot policies or Vision-Language-Action (VLA) models, it assists robots in action generation, causal reasoning, and embodied decision-making. Physical AI must answer two fundamental questions: how the world changes next, and how the world reacts after entities perform actions. Putting the two together, the roadmap becomes clear: first build a simulatable virtual world, let machines learn within it, and then transfer the learned capabilities back to the real world.
From an academic perspective, this is also a key leap in the AI paradigm: upgrading from perceptual and generative AI ('understanding, generating images and text') to predicting physical causality and undertaking embodied decision-making. Many discussions tend to simplify world models into high-quality video generation, which is a significant cognitive bias—in industrial scenarios, simulability and verifiability are far more important than visual realism.
Wu Kan: Academician Guo’s definition, I have heard the most plain version in industrial scenarios. Factories don't care about the term 'world model'; they care about whether changing robot types requires stopping the line for three days, or whether new workpieces require veteran technicians to re-teach. If skills practiced in a virtual world can be directly moved to the production line, reducing downtime by one day translates to tangible profit. The translation of 'physical AI' in the workshop is flexibility, efficiency, and safety.
Q2:Why choose these three parties to build a consortium together?
Academician Guo, based on the industrial physical AI route you proposed, why did you choose to partner with Estun and Freedo to build this consortium?
Academician Guo Song: Industrial physical AI cannot remain limited to papers and simulation demos. A complete closed-loop requires the superposition of three types of capabilities: theoretical frameworks and evaluation systems, real robot bodies and production line verification, and physically credible simulation environments. In our tripartite consortium, HKUST is responsible for defining the entire technical paradigm, world model algorithms, and acceptance standards; Estun, as a leading domestic industrial robot manufacturer, possesses mass-produced bodies, control links, and vast frontline process scenarios, serving as the carrier for technology implementation; Freedo has long focused on industrial simulation and spatial data infrastructure, capable of building physically aligned virtual training grounds. The combination of the three allows bridging the closed-loop of 'algorithm R&D → simulation pre-training → real-machine iterative verification,' validating the industrial physical AI route we proposed.
He Wenwu: Our positioning is the simulation infrastructure support within this technical closed-loop. The industrial world model paradigm proposed by HKUST has high requirements for the physical precision and interactive simulability of simulation environments, which aligns with our long-term accumulation. For world models to enter factories, model algorithms alone are insufficient; there must be real-world data, physically credible simulation environments, and real production lines and robots for verification. These three elements belong precisely to our three companies respectively. Model capabilities can be purchased, but capabilities precipitated from industry data and business closed-loops cannot be bought.
Q3:When will physical AI experience explosive development?
Following Academician Guo’s judgment, netizens asked a sharper question. Jensen Huang stated at CES 2026 that the ChatGPT moment for physical AI has arrived, proposing three major technical pillars for physical AI—the Newton physics engine, Cosmos world foundation model, and GPU + LPU hybrid computing architecture—and predicted that everything moving will eventually become a robot. Academician Guo, what is your view on this timeline and judgment?
Academician Guo Song: Let me first respond to this judgment. I agree with the general direction that physical AI is entering an industrial explosion phase, but we must distinguish between the 'concept inflection point' and the 'industrial implementation inflection point.' ChatGPT-style consumer-level explosions rely on standardized inputs, low error tolerance, and scaled C-end users; industrial physical AI is exactly the opposite, where every production line is a unique constraint system with extremely high thresholds for safety, processes, and compliance. The three pillars proposed by Jensen Huang represent the route of computing power and simulation infrastructure; however, the core bottleneck of industrial world models lies not in rendering compute power, but in physical alignment, uncertainty quantification, and real-machine migration—this is also the main focus of our HKUST team. Many people expect large-scale adoption of physical AI within one or two years; academic capability verification will be fast, but robust industrial implementation is inevitably gradual.
Q4:How do you view the value of physical interaction data and the complementary relationship between simulation data and real-machine data?
One of the most discussed topics among some observers is data. Multiple authors have mentioned that physical interaction data is the scarcest asset in the world, with real-machine collection speed matching the ticking of a wall clock one-to-one. Academician Guo, from an academic research perspective, how do you view the value of physical interaction data and the complementary relationship between simulation data and real-machine data?
Academician Guo Song: Physical interaction data is indeed a scarce asset, but we must not fall into the misconception that 'only real-machine data is useful.' Real-machine data is costly, difficult to collect in dangerous conditions, and severely lacks long-tail samples; the value of simulation data lies in low-cost generation of massive diverse interaction samples. However, simulation data has a fatal prerequisite: it must undergo physical alignment. Unaligned simulation data fed to models only causes distribution shift; models perform well in virtual environments but fail immediately on real machines. This is why we set action consistency and physical credibility as pre-conditions for the industrialization of world models. Simulation does not replace real machines but constructs a closed-loop of 'simulation pre-training — real-machine calibration — failure case feedback to simulation.' Within this loop, real-machine data serves as a calibration anchor, not the entirety of training material.
Wu Kan: The logic applies to Estun’s process data as well. Currently, these parameters, after anonymization and abstraction, become process benchmarks for the simulation world. New processes are first compared against benchmarks, and significant deviations directly expose problems. Data assetization is just beginning for Estun; the scenario library of the consortium makes this operationally feasible for the first time.
He Wenwu: Freedo has long focused on machine understanding of the real physical world, with a core high-precision calibratable simulation infrastructure. We undertake HKUST’s world model algorithm requirements, interface with Estun’s robot processes and real-machine data, complete physical parameter alignment for dynamics, materials, etc., supporting this data closed-loop to give simulation data the credibility required for industrial physical AI training.
Embodied Intelligence Driven by World Models | What Should Be the Core Technologies
PAL Physical Alignment Learning Framework: Connecting Prediction, Learning, and Decision-Making
Q5:How does physical alignment technology support industrial implementation?
Academician Guo just proposed four requirements and mentioned the team’s pioneering physical alignment technology. How exactly does this framework operate? Why is it considered the methodology for industrial implementation?
Academician Guo Song: Let me provide some background first. I am the Director of the Center for Physical Artificial Intelligence at HKUST and Chair Professor in the Department of Computer Science and Engineering. My research focuses on artificial intelligence, edge intelligence, and distributed computing. I am also a Fellow of the Royal Society of Canada, a Fellow of the Canadian Academy of Engineering, a Foreign Member of the Academia Europaea, and an IEEE Fellow. The HKUST Center for Physical AI has a cross-disciplinary core research team, including PhD students and postdocs from top Chinese universities such as Tsinghua University, Peking University, Nanjing University, USTC, and Fudan University. Back to the main topic: With robot bodies and training grounds ready, how capabilities land on real machines is today’s true hurdle. There are three obstacles ahead: First, the deviation in virtual-real interaction, where deviations in contact, friction, and deformation alter action outcomes; Second, the high cost of repetitive practice, where trial-and-error, reset, and real-machine verification continuously consume time and equipment; Third, difficulty in coherent execution of multi-step goals, where current actions need to connect to the target state of the next stage. To address these three hurdles, we proposed the PAL Physical Alignment Learning framework, progressing through three layers.
Academician Guo Song: L1 is Predictive Physical Cognition, with the core work being IRASim, published at ICCV 2025. It predicts physical changes caused by actions, performing fine-grained predictions on contact, slip, and deformation, rather than just outputting images; comparing model predictions with real recordings verifies whether contact positions, slip trends, and force/deformation are aligned. NVIDIA Cosmos was the earliest to use IRASim as a standard comparison baseline for action-conditioned prediction. L2 is Adaptive Behavior Learning, with the core work being WMPO, published at ICLR 2026. It completes reinforcement learning trial-and-error within the world model, placing the majority of trials in simulation, using real machines only for calibration and inspection; DreamDojo cites WMPO, discussing expanding action coverage through policy trial-and-error. L3 is Goal-Driven Task Decision-Making, with the core work being HALO, published at ICML 2026. It unifies textual reasoning, sub-goals, and continuous actions, using embodied chain-of-thought to break down long tasks into executable sub-goals; NVIDIA Cosmos 3 cites HALO in related work on unified embodied models.
We use industrial insertion as a typical task case for PAL to illustrate how this framework connects prediction, learning, and decision-making. Inserting a connector accurately into an interface breaks the task into four steps: positioning, alignment, insertion, and inspection; the world model first predicts contact consequences, identifies anomalies when actions deviate and encounter resistance, proactively retreats to realign, then aligns and inserts, finally completing the closed-loop through interaction and evaluation feedback. Completing these four steps demonstrates the entire process of L1 prediction, L2 learning, and L3 decision-making connected on a single production line.
Q6:What are the core technical requirements for industrial-grade world models?
In the first part, we discussed the technical core. Academician Guo has repeatedly cited embodied intelligence driven by world models as the main line in academic reports. Your judgment is that truly useful robotic world models cannot merely pursue visually plausible generated results. Please elaborate on the core technical requirements for industrial-grade world models.
Academician Guo Song: Four requirements. First is Action Consistency: the world evolution predicted by the model must align with the results of the robot’s actual actions, and action parameters must transfer directly from the virtual environment to the real machine. Second is Physical Credibility: physical laws such as gravity, friction, collision, and materials must hold up within the model; in industrial scenarios, this step is a prerequisite for entering the control closed-loop. Third is Uncertainty Awareness: the model must know in which scenarios it is unreliable, down-weighting confidence where necessary and triggering re-reasoning when appropriate. Fourth is Efficient Prediction: inference costs must be reduced so that world models can transform from expensive lab toys into deployable industrial components. Our team’s pioneering physical alignment technology upgrades generative video models into interactive world models, marking AI’s transition from observing the world to actively changing it. But if any of these four requirements is missing, it cannot be used on the industrial floor.
I want to add here: many studies treat world models as independent generative models, evaluating them solely on generated image quality; however, for embodied intelligence-oriented world models, the evaluation metrics should not be aesthetics, but rather action transfer success rate, prediction error bounds, and uncertainty calibration accuracy. This evaluation system is precisely the key work HKUST’s team is building within the consortium. Without corresponding evaluation standards, we cannot judge whether a world model can go onto the production line.
Q7:Why did industrial embodied intelligence begin to consolidate implementation in 2026?
Before expanding on the four requirements, let’s ask a timing question. The concept of world models is not new; why specifically in 2026 did industrial embodied intelligence begin to consolidate implementation?
Academician Guo Song: Three conditions matured simultaneously this year. The first condition is model capability: generative video models, after two years of iteration, can already produce visually credible world scenes, and physical alignment technology upgrades them into interactive world models, representing an academic inflection point. The second condition is infrastructure: adaptation of domestic computing power bases is complete, reducing the construction cost of training grounds by an order of magnitude. According to CAICT disclosures, over 70 training grounds have been built and put into use nationwide, transforming infrastructure from individual cases into an industry. The third condition is demand-side: the pressure for manufacturing flexibility has reached a critical point; robot manufacturers need capabilities that allow type changes without re-teaching, and factories need downtime compressed from weeks to days. Missing any one of these three conditions means the story doesn’t hold; in 2026, all three are present.
There is another academic layer: the maturity of VLA (Vision-Language-Action) models allowed us to find the fusion method between world models and robot decision chains. In earlier years, world models, robot control, and large vision models were three disconnected lines; now the academic community has bridged the interface frameworks, creating the basis for engineering implementation.
Wu Kan: Academician Guo spoke of macro conditions; I will supplement with the industrial side’s feeling. The demand for flexible transformation in manufacturing is becoming increasingly rigid; robots cannot always rely on manual teaching, which is a common pain point on the industrial end. Currently, downstream manufacturing universally moves towards multi-variety, small-batch, and rapid iteration. Industries we serve most frequently, such as NEVs, auto parts, and lithium batteries, have significantly accelerated product update rhythms. On real production lines, this encounters very practical difficulties: every time a type of workpiece changes, senior engineers often need to perform extensive teaching and calibration on-site, causing production lines to stop and wait, where downtime equals direct loss of capacity and revenue. Simultaneously, the industry faces the reality of a shortage of skilled process engineers; experienced veterans are scarce, and the mode of highly relying on human tuning is difficult to support large-scale flexible upgrades in manufacturing. In the past, body hardware had already reached high standards, but 'strong hardware, weak intelligent adaptation' is a universal shortcoming in the industry. Relying solely on local controller logic makes it hard to handle disturbances caused by working condition shifts and fixture changes. It is precisely these real pressures on the frontline that drove us to partner with HKUST and Freedo to truly implement world models and virtual training grounds on production lines, shifting extensive on-site teaching and tuning work to the virtual environment, fundamentally alleviating the contradictions of difficult line changeovers, long debugging cycles, and talent gaps.
Q8:What is the independent value of industrial physical AI?
Many discussions categorize world models under the narrative of general artificial intelligence, rarely discussing the independent value of industrial physical AI separately. Academician Guo, what is your view on this perspective?
Academician Guo Song: This is a crucial distinction. General world models pursue open-world generalization, encouraging exploration and tolerating trial-and-error; industrial physical AI is constrained intelligence, whose goal is not to 'understand everything,' but to achieve predictability, auditability, and accountability within given production line boundaries. While the underlying technologies share origins, their evaluation systems, safety boundaries, and optimization objectives are completely different. General models are evaluated openly; industrial world models must bind all capabilities to process indicators and safety indicators. This is the independent value of industrial physical AI. We do not oppose exploring general routes, but directly applying general world models to factories will inevitably lead to safety and reliability issues. This is also the research positioning of the consortium: not following the general model route, but prioritizing the establishment of a world model paradigm under industrial constraints.
He Wenwu: My view is straightforward. The biggest difference between industrial and consumer scenarios is the margin for error. Consumer products can pursue surprise; industrial systems must pursue reliability. Judging an industrial-grade spatial intelligence system, we look at only three things: Computable (objects in the model know the relationships between position, structure, material, function, and environment); Verifiable (output results can be tested against real data); Decidable (final outputs support specific business choices). Technical demos prove capability exists; industrial foundations prove business reliance. The independent value of industrial physical AI stands on these three points.
Cooperation Point 1: Action Consistency
Q9:What are HKUST’s core research tasks regarding action consistency?
Entering specific cooperation points. The first cooperation point is action consistency. Academician Guo, what are the core research tasks undertaken by HKUST along the line of action consistency?
Academician Guo Song: Action consistency is essentially a problem of representation learning and action alignment. HKUST’s role is to establish the methodologies for representation learning, action alignment, and decision fusion in world models, then use evaluations to force models to meet standards. The Department of Computer Science and Engineering at HKUST has a complete research echelon in embodied intelligence, covering everything from representation learning to policy learning, from simulation to real-machine transfer, with specialized research groups for each link. This is our confidence in undertaking the two cooperation points. A virtual environment with consistent actions is the starting point for policy training; no matter how well policies are trained in the virtual world, if action parameters don’t match the real world, everything is zero.
In terms of evaluation, IRASim is the carrier for L1 Predictive Physical Cognition. We do not make indiscriminate predictions for the entire scene, but only predict physical changes caused by actions: whether contact positions, slip trends, forces, and deformations are consistent with reality. Along the line of physical alignment, we examine contact and slip consistency, and force and deformation consistency—these are the quantitative scales of action consistency at the physical level.
Wu Kan: Estun’s role in this link is very specific. Moving action parameters from the virtual environment to the real machine involves physical details such as servo response, reducer clearance, and load changes. Our control engineers and HKUST researchers are currently aligning these parameters one by one. A one-percent difference in the lab is fine; a one-percent difference on the production line means scrap. Feedback from the real world is the most honest and demanding.
From Estun’s perspective, action consistency ultimately lands at the controller and actuator level: strategies practiced well in the virtual world must be writable as instruction sequences directly executable by our control systems. Our cooperation with HKUST includes writing this alignment into the interface specifications of controller firmware, transforming 'simulation-to-real-machine transfer' from engineer-by-engineer manual adjustment into a reproducible process at the firmware level.
He Wenwu: Freedo provides simulation infrastructure support. We provide 3D and video-fused simulation representations, offering a foundational environment for action alignment, bridging the physical mapping between virtual space and real production lines, undertaking HKUST’s world model prediction and training requirements, interfacing with Estun’s real-machine process parameters, ensuring consistency in dynamics and contact characteristics between simulation and real machines, supporting the implementation of the 'simulation pre-training, real-machine calibration, case feedback' data closed-loop.
Q10:What is the relationship between world models and VLA?
Academician Guo, answering a basic question for tech-savvy netizens. Currently, Vision-Language-Action (VLA) models are very popular in robotics. What is the relationship between world models and VLA?
Academician Guo Song: It is a division of labor. VLA handles mapping perception and instructions to actions, answering 'what to do now'; world models handle predicting how the world changes after actions, answering 'what happens after doing it.' One manages the first half of decision-making, the other the second half. Both segments must be robust in industrial scenarios: VLA understands the semantics of the production line, world models deduce the consequences of actions, and uncertainty awareness acts as a safety net in between. In our dual-role framework for world models, the aspect acting as an internal prediction module within policies works embedded in the VLA system. The convergence of these two technical lines on the industrial floor is a shared discovery by academia and industry this year—different paths, same destination.
I will add a boundary: VLA excels at semantic understanding and task decomposition, but it does not inherently possess physical causality. Without world models for consequence deduction, the action sequences output by VLA may be physically infeasible or pose collision risks. Only the combination of the two constitutes a complete embodied decision-making chain.
Wu Kan: On the product side, our control base AI Agent OS and dedicated Juliet robot language are integrating VLA and world models into a unified technology stack. Past pain points in the robot industry included controllers and vision systems speaking different languages, with integration relying on engineers’ manual workshops. With a unified technology stack plus world model deduction capabilities, robots have the possibility for the first time to build application ecosystems on standard bases, like smartphones.
He Wenwu: For this chain, the simulation infrastructure is the carrier of physical causality. VLA generates action instructions, and world models deduce consequences; both require a physically credible virtual environment for verification. Our 3D and video-fused simulation representations can rehearse VLA-outputted actions in simulation space, pre-validating whether dynamics and contact interactions are reasonable, providing a reproducible physical test field for the entire decision-making chain.
Q11:What effects should this cooperation achieve?
Academician Guo Song: The effect is reducing the transfer loss of the trained robot brain from simulation to real machines to an engineering-usable level. HKUST labs are connected with Freedo’s simulation environments and Estun’s real-machine control links. Strategies developed by PhD students and researchers run first on virtual production lines, then calibrate on Estun’s real machines. The cycle from lab to factory can now be calculated in months for the first time. Previously, student strategy tuning took months from code to real-machine verification, with much time spent waiting for production line windows and repeated calibrations. Now, strategies validated on virtual production lines go directly to real-machine calibration, reducing transfer loss to engineering-usable levels. The speed of research iteration becomes the speed of industrial progress.
Wu Kan: For Estun, this effect equates to changing the verification cycle of new strategies from 'waiting for production line windows' to 'runnable anytime.' Previously, strategy iteration debugging required accommodating factory production plans for real-machine testing; wanting to run a full verification often meant queuing for production line shutdown intervals, severely constraining iteration efficiency.
Our control engineers can directly perform real-machine reviews against parameters already calibrated in the training ground. Extensive boundary conditions and action logic trial-and-errors are completed entirely in the virtual training ground, no longer repeatedly occupying production lines for errors. Real machines are only used for final confirmation and calibration, no longer bearing heavy trial-and-error workloads. This is also the period customers most hope to shorten during existing production line transformations, minimizing debugging interference with actual production.
He Wenwu: We are responsible for solidifying the physical consistency base of the virtual training ground. Relying on 3D and video-fused simulation representations, we continuously align dynamics and contact interaction characteristics between virtual scenes and real production lines, ensuring that HKUST’s strategies and Estun’s control logic trial-and-error results in simulation are transferable, supporting the stable operation of this high-speed iteration closed-loop.
Cooperation Point 2: Physical Credibility
Q12:How is 'physically credible' measured?
The second cooperation point is physical credibility. Academician Guo, why is physical credibility a non-negotiable bottom line for industrial world models? How does academia measure 'physical credibility'?
Academician Guo Song: Physical credibility essentially requires the model to adhere to known physical constraints, rather than simply fitting video distributions. Many generative world models are statistical fits; they look visually coherent, but acceleration, collision responses, and load transfers are wrong. Once connected to a robot control closed-loop, such false physical predictions directly cause safety accidents. Measuring credibility cannot rely solely on visual effects; quantitative metrics are needed: dynamic error, collision response error, and material deformation prediction error. This part requires joint calibration between simulation environments and real robots, which is also the focus of tripartite collaboration.
To measure physical credibility, we do not look at visual aesthetics, but at three sets of quantitative indicators: whether contact and slip are consistent, whether force and deformation are consistent, and whether dynamic error, collision response error, and material deformation prediction error fall within acceptable bounds. Many generative models are merely statistical fits, visually coherent but physically incorrect; connecting them to control closed-loops causes safety accidents. These three sets of indicators require joint calibration between simulation environments and real robots.
Wu Kan: Whether the physical laws in the simulation world are accurate is ultimately voted on by the real production line. Estun’s robot production lines are themselves the best testing grounds—we first use real machines to run data and verify models, then push to customers. If simulation cadence, trajectories, and energy consumption don’t match the real production line, engineers will shut it down on the first day. The veto power of the production line is the hardest endorsement of physical credibility.
We hold vast amounts of real working condition and process data, from servo response to load elasticity changes; these are the links most prone to distortion in simulation yet most impacting reliability. Physical credibility is not determined by guessing simulation parameters, but calibrated bit by bit using real data accumulated on our production lines.
He Wenwu: Freedo provides physical simulation engines and real-world data, incorporating production line geometry, processes, and physical laws into the simulation world; Estun provides bodies, control interfaces, and real production lines, pulling simulation results back to reality for inspection. What we have done daily for the past decade-plus is hooking up real data and letting real feedback score the predicted results. Only such worlds qualify as training grounds for robots.
Q13:How exactly do the real-machine route and simulation route cycle? What are the engineering details?
How exactly do the real-machine route and simulation route cycle, and what are the engineering details?
He Wenwu: Real-machine data performs last-mile calibration, while scaled pre-training is completed in the simulation world. Bidirectional cycle: Simulation provides pre-trained policies, real machines provide calibration data and failure cases, failure cases feed back to simulation, becoming targets for the next round of training. Each turn of the cycle brings the simulation world closer to the real factory.
Wu Kan: From the perspective of the finished machine manufacturer, real-machine feedback has a hidden value: exposing simulation blind spots. Failure cases occurring on real machines are often scenarios nobody thought of in simulation—micro-vibrations of robot grippers under high loads, interference of production line lighting changes on visual recognition. These cases feed back to simulation, forcing the simulation world to grow abilities it previously lacked. Our process engineers and Freedo’s simulation teams are currently filling these blind spots one by one.
Cooperation Point 3: Uncertainty Awareness
Q14:Why is uncertainty awareness equally critical for factories?
The third cooperation point is uncertainty awareness. This sounds the most academic; why is it equally critical for factories? Academician Guo, please take the lead.
Academician Guo Song: The margin for error on the industrial floor is small; if a robot misinterprets an action topology as safe redundancy, the consequence occurs in milliseconds. Uncertainty awareness requires models to output confidence, proactively down-weighting situations exceeding the training distribution, triggering re-reasoning or retreating to verified deterministic controls. Our goal is to embed reasoning and verification into the policy learning process, allowing skill accumulation to accompany auditable reliability boundaries.
Many AI models only output a certain prediction, not telling users 'whether this prediction is reliable.' In industrial scenarios, this is a fatal flaw. Uncertainty awareness is not a nice-to-have, but a safety fallback mechanism. When models encounter out-of-distribution working conditions—such as workpiece deformation, fixture offset, foreign object intrusion—after confidence drops, the robot does not continue executing high-risk actions but switches to conservative control strategies. This is also a key focus of our research group: natively integrating uncertainty quantification modules into world models, rather than externally attaching classifiers afterwards.
Adaptive behavior learning WMPO plays a role here. Facing VLAs, it performs on-policy reinforcement learning within the world model, iteratively optimizing imagined trajectories via GRPO based on reward signals, where reward 0 represents failure and 1 represents task completion. The majority of trial-and-error is placed in virtual environments, with real machines used only for calibration and inspection, greatly improving sample efficiency and reducing real-machine wear. Industrial insertion is an intuitive example: ordinary strategies continue pressing down upon collision, leading to insertion failure; WMPO-trained strategies automatically lift and realign upon collision, completing the insertion. This is an emergent behavior of self-recovery after collision.
Occluded conditions have another layer of key evidence coming from the robot’s bodily feedback. Taking thin towel grasping as an example, vision cannot see the contact state, but gripper torque feedback can confirm physical contact: when torque continuously deviates from the reference value beyond a threshold, contact is judged to have occurred; after closing the gripper, a light lift verifies that only if the towel moves with the gripper is grasping confirmed successful. Without torque feedback, the same initial observation would result in an empty grab, leaving the towel on the table. Using bodily feedback to confirm grasping is an engineering means to avoid visual occlusion misjudgments.
Wu Kan: This item has very specific correspondences on the production line. Factory working conditions never perfectly match training data: workpieces have machining tolerances, fixtures wear and age, material loading positions shift, and workshop lighting conditions change constantly, with various disturbances emerging endlessly. Many boundary cases are difficult to exhaustively cover in training datasets. Previously, encountering these uncertainties relied solely on the on-site experience of veteran technicians to adjust and remedy, resolving various anomalies manually. However, the supply of skilled process personnel is limited, and relying on human experience to bear risks is difficult to replicate and promote on a large scale.
Our cooperation with HKUST deeply integrates the confidence mechanism into the robot’s underlying control logic. When the model judges 'I'm unsure this time,' meaning the current working condition has deviated from the training distribution, the robot does not stubbornly execute high-risk actions but proactively decelerates, reduces action amplitude, or retreats to verified deterministic programs, alerting for human intervention if necessary. Transforming human experiential judgment into built-in robot safety decisions safeguards the safety bottom line of production line operations while maintaining production efficiency.
He Wenwu: Applied to products, this is the value of the intelligent evaluation system. Training grounds organize complete capability chains by generation, understanding, deduction, and decision-making, with evaluation stages providing credibility ratings for each type of scenario. When customers procure robot solutions, they receive an acceptance report with confidence levels. A demo video cannot prove reliability; procurement decisions for industrial systems require such evidence.
Cooperation Point 4: Efficient Prediction
Q15:How to solve the dilemma of excessive inference overhead in world models?
The fourth cooperation point is efficient prediction. For world models to enter production lines, computing power cost is a hard threshold. Academician Guo, how does academia solve the dilemma of excessive inference overhead in world models?
Academician Guo Song: For world models to enter the interior of robot policies, inference latency and computing overhead must match the control frequency. Our direction is model compression, lightweight prediction, and edge deployment, transforming world models from large cloud models into components runnable at the edge of production lines.
A core idea is decoupling: there is no need for continuous high-resolution generation of the entire scene, only high-precision physical deduction for local areas interacting with the robot; irrelevant areas use low-precision representations. Combined with sparse prediction, temporal caching, and task-driven dynamic precision scheduling, inference latency is compressed within the robot control cycle. Purely stacking computing power offers no way out; the computing budget of industrial controllers is rigid. Lightweight, task-aware prediction architectures are the path to implementation.
Wu Kan: Production lines have simple and brutal requirements for computing power: sufficient, stable, no disconnections. Training can be run slowly on local or cloud computing power, but inference and control must be completed in real-time at the edge of the production line. Estun’s big-small brain control systems and servo systems are themselves edge computing nodes. Sinking world model inference to this level minimizes integration costs and is the prerequisite for truly running models at the edge of production lines—computing power is not built in the cloud, but grown on the machine itself.
He Wenwu: Freedo cooperates on the domestic computing power side. The Zhengrong model has completed domestic computing power adaptation for autonomous 3D calculation, rendering, and physical simulation engines; the Yuanzun all-in-one machine places training and evaluation computing power locally at customer sites. Model capability is one aspect; controllable computing power is another; both must be self-reliant.
Cognitive Gap 1: Quality Standards for Training Grounds
Q16:What kind of simulation environment qualifies as a training ground for industrial robots?
Simulation tools are increasing, and everyone is building virtual environments. Academician Guo, what kind of simulation environment qualifies as a training ground for industrial robots? Who should define the quality standards for training grounds?
Academician Guo Song: This is one of the tasks in the second phase of the consortium. Our proposition is that training ground quality standards must simultaneously consider four aspects: action consistency, physical credibility, uncertainty awareness, and efficient prediction—all indispensable—plus industrial acceptance. Beyond public leaderboards, industrial acceptance must become an independent evaluation tier.
The industry currently lacks unified evaluation benchmarks oriented towards embodied intelligence training effectiveness, not simulation software. Traditional simulation software only validates geometry and rendering; training grounds are used to train robot policies, and the evaluation object is 'robot real-machine performance after training,' not the simulation screen itself. This is a completely different evaluation paradigm.
The consortium’s training ground, officially named WorldFoundry, is a physical-consistency embodied simulation and training workshop, led by HKUST, with Freedo Technology and Estun as strategic partners. It targets robot training and evaluation, automatically constructing and continuously calibrating executable virtual worlds, divided into three major modules: Asset Generation and Verification, generating interactive digital assets and verifying their physical behaviors; Executable Task Worlds, constructing scenes and generating interaction data according to robot and task requirements; Robot Evaluation and Calibration, assessing task performance, analyzing simulation-to-reality discrepancies, and reverse-calibrating environments.
Around the same batch of real industrial tasks, PAL penetrates three-layer evaluation: L1 verifies whether action consequences are consistent, comparing predicted and measured contact, slip, and deformation; L2 verifies whether practice benefits translate to real machines, looking at success rates, work cadence, and human intervention; L3 verifies whether complete tasks can be stably accomplished, examining multi-step execution and recovery after state changes. Public leaderboards focus on screen generation effects; WorldFoundry’s industrial acceptance focuses on actual real-machine performance after simulation training, ignoring simulation screen rendering quality.
This training ground drives the evolution of industrial robots from 1.0 to 2.0. Phase 1.0 sees equipment and training capabilities ready; Phase 2.0 continuously improves around the same real industrial tasks: on-site difficulties are addressed by HKUST’s original algorithms combined with Freedo’s scenario training, practiced in virtual worlds, then verified on Estun’s real-machine bodies. After verification, deployment transforms real work experience into sustainably improvable product capabilities. Virtual practice, real-machine verification, and post-verification deployment are precisely the implementation of the three-party division of labor on the training ground.
Q17:What is the relationship between the training ground quality standards proposed by the consortium and national standards?
The MIIT embodied intelligence standard system has already released 45 sub-categories and 100 standards, and an industry alliance has been established. What is the relationship between the training ground quality standards proposed by the consortium and national standards?
Academician Guo Song: They are complementary and divided. National standards solve general problems: terminology, classification, safety baselines, and data formats, giving the entire industry a common language. Consortium training ground quality standards solve specific problems on the industrial floor: what constitutes action consistency, what constitutes physical credibility, how to calibrate confidence, and how to write acceptance reports. Our plan is to distill the acceptance methods run on real production lines by the consortium into standard drafts and contribute them to the industry. Standards are not written behind closed doors but polished on production lines. Whoever runs it through first has the responsibility to clarify the methods, helping successors avoid detours. China’s advantage lies precisely in scenario density; most of the world’s 70+ training grounds are in China, making the testbed for standards right here. We aim to mature the methodology of industrial acceptance in China first, then push it internationally, occupying an ecological niche in the landscape of world model evaluation.
Why Are the Three Companies Complementary?
Q18:What irreplaceable advantages does each company possess?
Having covered the four cooperation points. Please each describe your irreplaceable part in one sentence, starting with HKUST.
Academician Guo Song: HKUST provides the theoretical main line, evaluation methods, and talent for world models and embodied intelligence, answering where technology should go.
Wu Kan: Estun provides robot bodies, control bases, real production lines, and factory customers, answering where practiced capabilities undergo real testing.
He Wenwu: Freedo provides years of accumulated real-world data, 3D simulation infrastructure, and computing power adaptation capabilities, answering how virtual worlds are built to industrial-grade precision.
Q19:How are the technical routes of world models categorized?
There has been an influential debate among some observers. Some authors divide world models into the Spatial Geometry School and the Interactive Simulation School, with Atlas representing the former and Google DeepMind’s Genie 3 representing the latter. Academician Guo, how do you classify this?
Academician Guo Song: This classification framework hits the mark. The Spatial Geometry School builds the structure of the world correctly; the Interactive Simulation School evolves the changes of the world correctly. The answer for industrial scenarios is clear: both must be combined. Without geometric structure, robots don’t know where they are; without dynamic interaction, robots don’t know what happens next. Atlas has achieved excellence in rendering and reconstruction; Genie 3 has achieved excellence in real-time interaction. Truly complete virtual worlds require combining both.
This is also why our consortium chose the 3D + Video Fusion route: geometric structures handle localization perception, while physical interaction engines handle dynamic deduction. Single routes have shortcomings. Many academic teams delve deeply into one school each, but industrial implementation requires fusion. HKUST Center for Physical AI’s frontier exploration of embodied intelligence driven by world models, Estun’s deep accumulation in R&D and manufacturing of industrial robot bodies, and Freedo’s simulation environments form a perfect closed-loop. Results from body research need virtual environments to accelerate training; results from virtual environments need bodies for landing verification. The two routes are functionally isomorphic on the industrial floor, taking different paths to the same destination, ultimately arriving at interactivity, simulability, and verifiability.
Comparison with the World’s Most Advanced Research
Q20:What is your view on Atlas?
Discussing consortium advantages, international benchmark Fei-Fei Li’s team Atlas cannot be avoided. Academician Guo, what is your view on Atlas?
Academician Guo Song: Atlas is an impressive engineering achievement. Professor Fei-Fei Li categorizes world models into renderers, simulators, and planners. Atlas has already become outstanding in rendering and reconstruction, rebuilding geometrically credible worlds from a few photos, with reconstruction accuracy surpassing specialized models. Current judgments on physical AI by academia and industry are highly consistent, essentially converging. Rebuilding geometrically credible worlds from a few photos has the limit of rapid generalization; building 3D and video-fused models from real-world data has the limit of industrial precision. The existence of these two routes indicates that the technical landscape of world models is much broader than currently seen, with sufficient academic space to accommodate multiple explorations.
He Wenwu: I also hold great respect for Professor Fei-Fei Li. The competition in world models: the first half is generation, the second half is understanding, deduction, and decision-making. We decompose using the same framework as World Labs: Zhengrong corresponds to generation, physical simulation corresponds to simulation, and decision models correspond to planning; functions are isomorphic. The difference lies in the route: Atlas rebuilds worlds from photos; Freedo starts from real-world data, fusing 3D and video, embedding physical simulation capabilities within the model system. Generating an image of a world and understanding a running world are two different capabilities.
Wu Kan: From the perspective of an industrial robot finished machine manufacturer, the general reconstruction capability demonstrated by Atlas is brilliant and highly inspiring. However, the dimensional needs of industrial scenarios are more diverse: industrial scenarios require not only environmental appearance reconstruction but also physical attributes like friction, load, and contact deformation, which are harder to obtain directly from photos. Atlas excels at quickly building scenes from images, while we emphasize closed-loop verification of real process data and robot hardware. Both routes have strengths; academically, flowers can bloom in variety; but landing on manufacturing production lines, stability, process adaptability, and delivery costs run by real machines ultimately matter most.
Priorities of Technical Routes
Q21:How do you view the priority between industrial and humanoid robots?
Humanoid robots have greater buzz than industrial robots, and many views default to assuming the main battlefield of physical AI is humanoid. Academician Guo, from an academic route perspective, how do you view the priority between the two?
Academician Guo Song: I agree with the implementation order of 'industry first, humanoid later,' but the reason goes beyond commercial payment. From a research perspective, industrial robot task spaces are relatively structured, with clear constraints and easier-to-define safety boundaries, making them the best testing ground for physical AI. The world models, uncertainty awareness, and evaluation systems polished in industrial scenarios are themselves the infrastructure future humanoid robots will need. Humanoid robots face unstructured open environments with more complex sources of uncertainty and higher research difficulty. Industrial scenarios are a necessary stage for us to accumulate methodologies and polish evaluation systems, not the endpoint. Once the capability base is bridged, humanoids are just one type of embodied carrier on this base.
Wu Kan: My answer is pragmatic: industry first, humanoid later. Industrial robots have an existing market today, with clear paying entities and rigid acceptance standards. Humanoid robots are an exciting direction, and Estun is watching, but large-scale entry of humanoids into factories still requires crossing hurdles of cost, reliability, and safety norms. Solving the flexibility problems of industrial robots first, solving the safety problems of human-machine collaboration, and establishing world models and training grounds firmly on the industrial floor means that when humanoid robots enter factories, they can step directly onto already-built training facilities and evaluation systems.
He Wenwu: This capability chain of training grounds does not discriminate by robot morphology. Robotic arms can use it, AGVs can use it, and future humanoid robots can also use it. Morphologies change, but physical laws and spatial relationships do not. We build the unchanging things solidly, and leave the changing things to time.
Visibility of Domestic World Models
Q22:How should domestic teams establish their own voice?
Domestic world models have weaker visibility on international public leaderboards. Academician Guo, how should we view leaderboard evaluations, and how do domestic teams establish their own voice?
Academician Guo Song: Public leaderboards, such as WorldRoamBench and WorldMark, have their value, suitable for horizontal comparison of general capabilities. But leaderboards themselves have limitations: most focus on open-world generation and visual reconstruction effects, lacking industrial constraint indicators. So our approach is dual-track parallelism: actively participate in international public leaderboards to prove basic capabilities; simultaneously independently build evaluation systems oriented towards industrial scenarios, namely the industrial acceptance specifications the consortium is advancing. Leaderboards prove academic potential; production line acceptance proves industrial value. Voice is not determined solely by leaderboard rankings, but established by reproducible, verifiable industrial effects.
Wu Kan: Adding from an industrial perspective, visibility is ultimately realized through customer orders. Leaderboards are mostly horizontal comparisons at the laboratory level; for industrial products, projects actually landed and rigorously verified on production lines are the hard currency. Consortium solutions will prioritize polishing and iteration on Estun’s internal smart production lines, realizing capabilities like action transfer, physical deduction, and uncertainty handling on real machines; once mature, leveraging our complete channel system of 75 global service points, overseas subsidiaries, and the Poland manufacturing base, we will export alongside robot finished machine solutions. Actual project acceptance sheets and overseas customer implementation cases are more convincing than any leaderboard ranking.
He Wenwu: Leaderboards can prove the upper limit of algorithms, but the industrial track competes on the lower limit of engineering implementation. Many models shine in public evaluations but easily fail when facing the diverse working conditions and complex process constraints of real factories. For the consortium, we do not reject participating in various international evaluations to show basic capabilities, but our true voice comes from precipitated real-scenario assets, simulation bases repeatedly calibrated on production lines, and replicable delivery solution packages. Leaderboards are business cards; standing firm requires one replicable industrial project after another.
Strategic Significance
Q23:Will the consortium become a talent cultivation platform?
How is HKUST’s research echelon in embodied intelligence constructed? Will the consortium become a talent cultivation platform?
Academician Guo Song: Yes, and from the beginning. HKUST has a complete research echelon in embodied intelligence, covering everything from representation learning to policy learning, from simulation to real-machine transfer, with specialized research groups for each link. The consortium provides this echelon with conditions previously unavailable: real production line data, industrial-grade simulation environments, and mass-producible robot bodies. Student research faces real industrial problems during their studies for the first time. We plan to jointly offer courses and co-build practical training projects, allowing PhD topics to grow directly on the consortium’s tackling list. Cultivated talents will understand both world model theory and factory physical constraints; such people will be extremely scarce in the next decade.
This is also our core lever for solving industry-academia-research integration: research topics stem from real pain points on production lines, and evaluation criteria include both paper metrics and production line verification metrics. Research is no longer self-closed; paper results must flow back to training grounds for verification. Students must understand physical constraints and safety boundaries from the start, rather than just chasing metrics on experimental datasets.
Wu Kan: It is undoubtedly beneficial to Estun’s internal engineering team, allowing valuable debugging experiences of veterans to be precipitated into standardized simulation benchmark cases, facilitating young engineers’ learning and practice in virtual environments, and breaking the long-standing talent gap in the industry. Its greater strategic value lies in serving as an open robot embodied intelligence ecosystem platform: we not only develop AI vertical solutions for flexible scenarios based on the platform but also open platform capabilities and supporting toolchains to broad ecosystem partners, supporting partners to independently develop various scenario applications based on the base, collectively serving more diverse flexible manufacturing needs.
He Wenwu: Freedo acts as the translator in the middle. We understand both the language of engineering delivery and the boundaries of model capabilities, knowing which production line problems can be modeled and which lack conditions. This translator role is key to the three parties meshing together.
Q24:How to bridge industry-academia-research cooperation and industrial demands?
University achievements in industry-academia-research often mismatch enterprise demands. How does the consortium break this deadlock?
Academician Guo Song: The key to breaking the deadlock is letting problems grow from production lines, not from literature. The first item on the consortium’s tackling list is the real pain point on Estun’s production line; our research methods are organized around this list, and students know which production line their topic is responsible for on their first day in the group. Papers and acceptance sheets use the same dataset, eliminating the 'two skins' phenomenon.
Wu Kan: Enterprises also need to change habits. Previously, enterprises seeking university cooperation tended to throw out minor improvements to mature products while keeping core technologies secret. Within the consortium, we open real working conditions and process data to research teams, with confidentiality agreements clarifying boundaries, and data used for tackling challenges. Openness buys research speed; the math adds up.
Q25:What is the industrial value of the cooperation?
The technical core is now clear. Finally, we ask the three experts to discuss the significance separately.
Academician Guo Song: I will speak to the significance for AI development. The past decade was the decade of perceptual intelligence, where AI learned identification and generation, and language models made AI expressive. The next decade is the decade of physical intelligence, where AI learns to act in the physical world, and robots give AI a body. Jensen Huang summarized AI evolution into four stages: Perceptual AI, Generative AI, Agentic AI, and Physical AI. This framework aligns with academic judgments: the next station of intelligence definitely lands in the physical world. World models are the cognitive base of this leap, allowing AI to simulate and deduce in virtual worlds instead of repeatedly trial-and-erroring in real environments, greatly reducing hardware wear and safety risks. Embodied intelligence driven by world models is the inevitable path from usable to daring-to-use and stably-used.
In the long run, our industrial world model paradigm will also provide a reusable research paradigm for the domestic embodied intelligence field: centered on physical alignment, uncertainty quantification, and real-machine transfer, rather than purely chasing generated screen effects.
Wu Kan: I will speak to the industrial significance. China’s industrial robot market is the largest globally, and domestic robots have completed the first stage of import substitution; the next stage is flexible manufacturing and going global. Difficulties in implementing flexible lines, long debugging cycles, and shortages of high-skilled technicians block many factories. Training grounds driven by world models move debugging from real production lines to virtual worlds; virtual debugging allows layout changes and robot deployment schemes to run through in virtual factories first, then land on real lines, compressing downtime from weeks to days. This is the key leap for the robot industry from selling single units to selling systems. Estun’s goal is to build a world brand for Chinese robots, and the consortium is one of the reasons for this confidence.
He Wenwu: I will speak to the significance for China. Physical AI is the best validation ground for autonomously controllable computing power and the best arena for real-world data. China possesses the world’s largest industrial robot market, the most complete manufacturing scenarios, and the richest engineer dividend. The definitional power of world models cannot rest solely in others' hands. The consortium connects world models, spatial intelligence bases, and robot bodies into a closed-loop, with every step undergoing inspection on real production lines. Five years from now, if someone asks who made the physical world computable, deducible, and decidable, we believe the answer will certainly include Chinese companies and Chinese teams.
In Closing
This in-depth interview originated from the strategic cooperation signing ceremony of HKUST, Estun (002747.SZ, 02715.HK), and Freedo Technology at the 26th CIIF on October 12. In the future, the three parties will take the industrial physical AI and world model paradigms discussed in this interview as the core main line, jointly building an industrial physical AI technology base, constructing the closed-loop chain of 'Algorithm R&D — Simulation Pre-training — Real-machine Iterative Verification,' cracking industrial bottlenecks such as virtual-real transfer, physical alignment, and uncertainty handling, driving embodied intelligence driven by world models from laboratory research to real production lines, and supporting flexible manufacturing transformation and the implementation of new productive forces.