ByteDance Robotics Veteran Tao Kong Joined Xiaomi, Raising the Stakes in Robot Foundation Models

Xiaomi reportedly recruited former ByteDance robotics leader Tao Kong in summer 2025, placing him in charge of a closely guarded robot foundation model team. The appointment matters because Xiaomi now has more than a research leader. It has an executive who spent six years building ByteDance's robotics program, along with direct access to Xiaomi's devices, vehicles, and factories.

The personnel move was reported by 36Kr on August 10, 2026, citing several independent sources familiar with both companies. Those sources said Kong brought several former ByteDance employees with him. They also placed Xiaomi's wider robotics organization at roughly 200 people, although Xiaomi has not publicly confirmed that figure.

This is not merely another senior AI hire. ByteDance built one of China's most visible general-purpose robot model programs under Kong, while Xiaomi has spent 2026 publishing its own models and training methods. The contest is now between two approaches to embodied AI: an internet platform developing robotics research from the model outward, and a hardware company building models around products, factories, and physical deployment.

What reportedly changed inside Xiaomi

Kong's arrival gives Xiaomi an experienced team builder at the point where its robotics work is moving from demonstrations toward a model platform.

The original report describes Kong as the head of Xiaomi's robot foundation model team. A foundation model is a broadly trained system that can be adapted to many tasks, rather than being designed for one fixed motion or machine.

The distinction is important. Traditional industrial robots repeat carefully programmed actions inside structured environments. Foundation model research aims to let a robot interpret language, understand changing scenes, and generate appropriate physical actions across different tasks.

Kong's public biography independently establishes the relevant background. On his personal research page, he says he served as ByteDance's director of robotics research from 2019 through 2025. He also says he established and led its robotics research team.

The page does not publicly name Xiaomi as his current employer. It does, however, invite researchers and engineers to work on robot foundation models and systems. That wording aligns with the reported scope of his new position, but it does not independently verify every organizational detail in the report.

According to 36Kr, Kong joined Xiaomi during summer 2025 and recruited several people who previously worked at ByteDance. The report also says his team uses a separate office. One source characterized the group as unusually secretive, even within Xiaomi's broader robotics organization.

Those descriptions should be treated cautiously. Neither company has published an appointment notice, an organizational chart, or a verified headcount for the group. Xiaomi has also not publicly addressed whether Kong directly manages every project released under its robotics research brand.

Still, Xiaomi's public research timeline gives the personnel report substantial context. The company introduced Xiaomi-Robotics-0 in February 2026, published its post-training pipeline in April, and announced two larger research projects in July. Those releases appeared after Kong reportedly joined.

That timing does not prove that Kong personally created each model. Research programs often begin long before publication, and papers usually represent work from many contributors. It does show that Xiaomi's robot model operation became publicly productive during the year following his reported arrival.

The most consequential part of the story is therefore not the secrecy surrounding one office. It is the convergence of an experienced research leader, a recruited team, a visible publication cadence, and a company with multiple paths to real-world deployment.

Why Xiaomi needs a robot foundation model team now

Xiaomi is trying to turn its hardware footprint into a training and deployment advantage before general-purpose robotics settles around a few dominant model stacks.

The company has worked on robots for years. It presented the CyberDog quadruped in 2021 and the CyberOne humanoid in 2022. Those systems attracted attention, but neither established Xiaomi as a supplier of general robot intelligence.

Its 2026 work makes a more ambitious claim. Xiaomi's robotics research site describes a program focused on general robots powered by scalable foundation models. Its releases emphasize model training, action generation, world modeling, and efficient adaptation to difficult manipulation tasks.

This direction reflects a wider change in robotics. A robot no longer needs only capable motors, joints, cameras, and control software. Teams increasingly want a reusable model that connects visual perception, language instructions, and physical action.

That class of system is often called a vision-language-action model, or VLA. It receives images and text, then generates actions that a robot can execute. In principle, one broadly trained VLA can be adapted to many machines and environments.

In practice, the task remains difficult. Robots must act in continuous physical space, where a small error can spill a drink, damage an object, or stop a production line. Internet-scale text and image data do not directly teach a machine how much force to apply when inserting an earbud into a case.

Xiaomi has assets that can help close that gap. It sells consumer electronics, operates connected-device platforms, manufactures electric vehicles, and has access to production environments. Each area offers possible deployment scenarios, although Xiaomi has not announced a unified commercial robot platform spanning all of them.

Factories are the most immediate testing ground. Repetitive manipulation tasks provide measurable success criteria, controlled safety boundaries, and economic reasons to automate. They can also generate the failure data needed to improve a model.

Homes present a larger market but a harder technical problem. Household objects vary widely, people move unpredictably, and acceptable failure rates are extremely low. A robot that succeeds nine times out of ten in a laboratory can still be unsuitable around children, pets, glassware, or hot appliances.

Vehicles create another possible bridge. Autonomous driving and robotics both depend on perception, planning, simulation, and decisions under uncertainty. Their sensors and action spaces differ, so models cannot simply transfer without adaptation. However, teams can reuse data infrastructure, training systems, and expertise in multimodal reasoning.

This is why an experienced foundation model leader matters now. Xiaomi must coordinate research across data collection, model architecture, post-training, evaluation, hardware control, and deployment. A collection of isolated demonstrations will not produce a reusable robot intelligence layer.

The reported team size of about 200 people also suggests Xiaomi sees robotics as a sustained program, not a small exploratory project. That number remains unconfirmed, but the company's publication schedule points to a substantial operation.

For Xiaomi, the opportunity is not only to build a humanoid robot. It is to create a model architecture that can serve different embodiments, meaning different robot bodies, sensors, and control systems. Success would give the company a common software layer for machines that currently require separate engineering.

Xiaomi and ByteDance are building from opposite ends

ByteDance brings large-model research and internet-scale AI infrastructure, while Xiaomi can connect model development to hardware, manufacturing, and eventual product deployment.

This is the primary contest created by Kong's move. Both companies can fund large research programs, recruit specialists, and train multimodal models. Their strongest assets are different.

ByteDance entered robotics from the model side. Its Seed research organization works across language, vision, speech, world models, and robotics. Under Kong's earlier leadership, its robot program pursued general models that could combine web knowledge with physical skills.

One example was GR-2, described as a generative video-language-action model for robot manipulation. The program later introduced GR-3, a four-billion-parameter model intended to follow language instructions, generalize across objects, and complete longer tasks.

ByteDance's GR-3 release says the model integrates vision-language understanding with action generation. The company demonstrated tasks involving flexible objects and dual-arm manipulation, areas where rigid programming often performs poorly.

Those demonstrations show ByteDance's research ambition. They do not establish broad commercial reliability. Public videos and selected laboratory tasks cannot reveal failure rates across thousands of homes or factories.

ByteDance also faces a deployment question. It owns major consumer platforms and cloud services, but it does not have Xiaomi's range of branded physical products. It can collaborate with robot manufacturers, license models, or build reference systems. Each route adds coordination between the model developer and the hardware owner.

Xiaomi approaches the problem from the other direction. It already designs devices and vehicles, manages suppliers, and operates manufacturing systems. Its challenge is building a general intelligence layer that matches the quality of specialist AI laboratories.

Recruiting Kong narrows that gap. He brings experience organizing a robot research team inside a large internet company. The reported arrival of former ByteDance colleagues also reduces the delay involved in assembling a group whose members already share methods and terminology.

Talent movement does not automatically transfer intellectual property. Employees remain bound by legal, contractual, and ethical limits concerning confidential information. There is no public evidence that Kong or the colleagues who followed him improperly transferred ByteDance technology.

What does transfer is experience. Researchers learn which data pipelines fail, which evaluations hide weaknesses, and which model designs become unstable during real-world control. That operational knowledge is difficult to reproduce from papers alone.

ByteDance still retains a visible robotics program. Its current research pages feature GR-3, dexterous manipulation hardware, and reinforcement-learning work. The company therefore remains a serious participant rather than a laboratory dismantled by one departure.

The pressure on ByteDance is strategic. It must show that a model-centered organization can turn strong research into repeatable deployment without owning the entire hardware stack. Xiaomi must prove the inverse: that hardware reach can produce better embodied intelligence, rather than merely more places to install an immature model.

Xiaomi's foundation model push now has measurable substance

Xiaomi's recent papers turn the hiring story into a testable technical strategy built around scale, world modeling, and efficient post-training.

Xiaomi-Robotics-0 was the program's first public foundation policy. A policy is the part of a robot model that maps its observations and instructions to actions.

The accompanying research paper describes pre-training on cross-embodiment robot trajectories and vision-language data. Cross-embodiment training combines data from different robot forms, with the goal of learning skills that do not belong to only one machine.

The researchers also focused on real-time execution. This matters because an action model that reasons too slowly cannot respond safely to a changing physical environment. Latency becomes part of intelligence when a robot must adjust its grip before an object slips.

In April, Xiaomi published a post-training pipeline for the model. Post-training adapts a broadly trained system to particular behaviors after its initial large-scale training. The company highlighted a precise earbud insertion task using 20 hours of task-specific post-training data.

That demonstration targets a genuine robotics problem. Small objects demand accurate positioning, force control, and recovery from slight misalignment. It is more informative than a robot walking through a prepared routine, although one successful task still does not establish general reliability.

July brought two additional projects. Xiaomi-Robotics-U0 focused on embodied world modeling, while Xiaomi-Robotics-1 emphasized scaling action models with much larger real-world datasets.

A world model predicts how an environment changes after an action. For a robot, that means estimating what will happen when it pushes, lifts, rotates, or releases an object. Better predictions can support planning before the machine commits to a physical movement.

The U0 paper argues that ordinary image and video generation models do not fully satisfy robotics requirements. Embodied systems need consistency across camera views, coherent geometry, and awareness of the robot's own physical constraints.

That is an important mechanism behind Xiaomi's strategy. Video generation can supply visual priors and simulated experiences, but a robot cannot treat the world as a sequence of attractive frames. Its predictions must remain compatible with three-dimensional space and executable actions.

Xiaomi-Robotics-1 pushes on data scale. Its researchers report pre-training on more than 100,000 hours of real-world manipulation trajectories collected using handheld interfaces. These interfaces allow people to generate demonstrations without operating every target robot directly.

The model paper presents a two-stage approach. Large-scale pre-training develops broad action-generation capabilities, then post-training adapts the model to complex tasks.

The reported dataset is significant because robot data is expensive. Text models can learn from documents produced for human readers. Robots need observations tied to actions, physical outcomes, and often specialized sensor records.

However, hours alone do not determine quality. A large dataset can contain repetitive movements, inconsistent demonstrations, narrow environments, or weak coverage of failures. The paper offers benchmarks and experiments, but independent reproduction will matter more than the headline number.

The sequence of releases reveals a coherent stack. Robotics-0 establishes the initial action model. U0 works on predicting embodied environments. Robotics-1 scales real-world behavior data. Post-training then adapts those general capabilities to precise tasks.

That coherence makes the reported appointment more consequential. Xiaomi appears to be constructing an integrated model program rather than publishing unrelated robotics experiments. Kong's experience at ByteDance fits the needs of such an organization.

It also changes how competitors must evaluate Xiaomi. The company is no longer competing only through robot hardware demonstrations. It is publishing architectures, methods, and models that researchers can inspect and challenge.

Open publication can help recruitment. Researchers often prefer organizations where they can release papers and build a visible body of work. Open models and training tools also create external feedback, even when the company retains its most valuable data or deployment systems.

Still, publication is not commercialization. Xiaomi has not identified a release date for a mass-market general-purpose robot based on these models. It has not disclosed field reliability, operating costs, safety incident rates, or customer adoption.

The technical strategy is now visible. The business and product strategy remains largely hidden.

The secrecy leaves three major questions unanswered

Xiaomi has shown credible components, but it has not yet shown that those components form a safe, economical, and dependable robot product.

The first uncertainty concerns leadership and attribution. The hiring report says Kong leads the foundation model group, but Xiaomi has not confirmed his title or responsibilities. Public papers list many researchers, and their work may have started before his arrival.

It would therefore be inaccurate to attribute Xiaomi's entire 2026 research portfolio to one executive. Kong's likely contribution is organizational as much as technical: selecting research priorities, recruiting teams, reviewing systems, and connecting experiments to deployment.

The second uncertainty concerns generalization. A model can perform well across laboratory tasks while failing when lighting changes, objects deform, or a person interrupts the sequence. Real environments contain edge cases that curated evaluations routinely miss.

Long tasks amplify these errors. A robot that has a 98 percent success rate on each individual step has only about an 82 percent chance of completing ten independent steps without failure. Actual errors are not always independent, but the simple calculation shows why polished demonstrations can mislead.

Recovery behavior matters as much as initial accuracy. A useful household or factory robot must detect a failed grasp, reassess the scene, and try another action without making the situation worse.

The third uncertainty is economic. Larger models and broader datasets demand compute, data collection, testing, and specialized hardware. A technically capable system can still fail if its operating cost exceeds the value of the labor it replaces or assists.

Factories offer controlled conditions, yet they already use mature automation. A foundation model must outperform fixed systems where flexibility has measurable value. Otherwise, customers may prefer cheaper equipment built for one task.

Consumer environments create the opposite problem. They contain abundant varied tasks, but buyers require affordability, safety, quiet operation, and minimal maintenance. The model cannot compensate for weak actuators, poor battery life, or unreliable sensors.

The reported organizational secrecy may protect competitive research and unpublished products. It also limits outside assessment. Xiaomi has not disclosed how its robotics group divides responsibility among model research, hardware, automotive systems, and manufacturing automation.

Secrecy can also create internal friction. Robot foundation models require close cooperation among data collectors, hardware engineers, safety specialists, and application teams. A separate office can protect focus, but excessive isolation can slow the feedback that physical systems require.

There is also a competitive response risk. ByteDance can accelerate its retained Seed robotics team, partner with hardware manufacturers, or make its models easier for developers to adopt. Other Chinese companies can recruit from both organizations and compete for the same limited pool of experienced researchers.

Google DeepMind, Tesla, Figure AI, Physical Intelligence, and several Chinese robotics companies are pursuing related goals through different hardware and model strategies. Xiaomi's access to manufacturing does not insulate it from advances elsewhere.

The current evidence supports a narrower conclusion. Xiaomi has assembled a serious research program and reportedly recruited a leader with relevant experience. It has not yet established that its models work reliably at commercial scale.

That distinction should guide how readers interpret future demonstrations. The important questions are not whether a robot completes one impressive task or whether a paper reports a larger dataset. The questions are how often the system fails, how it recovers, and whether its performance transfers to environments the developers did not curate.

For engineers and enterprise buyers, this is also a knowledge-management problem. Robot development generates experiments, failure videos, sensor logs, and decisions across many teams. Systems built for a searchable knowledge base can help teams connect that evidence, although documentation alone cannot solve model reliability.

What to watch after Tao Kong's reported move

Three signals will show whether Xiaomi is building a deployable robotics platform or an accomplished research laboratory with no near-term product.

The first signal is an official clarification of Kong's role and Xiaomi's organizational structure. Confirmation would establish whether he oversees the published foundation models, a future product stack, or a narrower research unit.

Leadership scope shapes execution. A research director can optimize papers and benchmarks, while a product leader must also own safety reviews, hardware constraints, deployment costs, and field support. Xiaomi's next hiring notices and contributor lists may reveal the practical boundaries before any formal announcement.

The second signal is independent technical validation. Researchers should look for reproduction of Xiaomi's results, public model evaluations, and performance on unfamiliar robots or environments.

Cross-embodiment transfer is especially important. If Robotics-1 can adapt efficiently to different machines and difficult tasks, Xiaomi's large pre-training investment gains leverage. If adaptation still requires extensive task-specific data, the platform becomes less general than its framing suggests.

Long-horizon evaluations deserve similar attention. A long-horizon task requires many dependent actions, such as clearing a table and placing several objects in their correct locations. Success rates should include full task completion, failure recovery, and safety interventions.

The 100,000-hour dataset also needs closer examination. Future papers could clarify its diversity, collection method, object distribution, environment coverage, and balance between successful and failed behavior. Those details will determine whether scale produced broad competence or mostly repeated familiar motions.

The third signal is a real deployment with operating metrics. A Xiaomi factory is the most plausible first venue because the company controls the environment and can measure productivity. A limited pilot would reveal more than another humanoid demonstration.

Useful metrics would include completed cycles, human interventions, mean time between failures, task changeover time, and the cost per successful operation. Xiaomi does not need to publish sensitive internal economics, but some evidence of sustained operation would support its broader claims.

A deployment in a vehicle plant would also test the company's strategic advantage. Xiaomi could connect robot learning with manufacturing workflows, sensor infrastructure, and quality-control systems already under its control.

A consumer launch would set a higher bar. Buyers would need clear safety boundaries, useful tasks, local or cloud processing policies, repair options, and predictable behavior around people. No public evidence currently indicates that such a launch is imminent.

These signals can strengthen or weaken the central judgment. Formal confirmation of Kong's broad authority would strengthen the view that Xiaomi is consolidating robotics around a foundation model strategy. Independent transfer results would show whether the technical stack is genuinely reusable.

A sustained factory deployment would provide the strongest evidence. It would connect research scale to physical value, which remains the unresolved problem across embodied AI.

The opposite outcomes would matter just as much. Continued silence about leadership, evaluations limited to curated tasks, or repeated papers without deployment would suggest that Xiaomi remains earlier in the transition than its publication pace implies.

Kong's reported move from ByteDance should therefore be read as the opening of a more demanding phase. Recruiting a respected robotics leader can accelerate research, attract colleagues, and sharpen priorities. It cannot remove the physical world's unforgiving requirements.

For developers, the next question is whether Xiaomi releases models, training code, or interfaces that can be tested outside its own laboratories. For enterprise buyers, the question is whether deployment data supports a credible return from flexible automation.

For everyone else, the clearest test is simple: watch what the robots do after the camera keeps recording. Xiaomi has assembled the people, data strategy, and hardware access needed to compete. Now it must show that those advantages survive contact with unstructured reality.