Inside TeleXperience, robot learning becomes a data loop

MORPHI KINO's 15-minute household run makes embodied AI's central problem visible: a robot must keep learning from execution, rather than merely complete a polished demonstration. According to geekpark, the robot handled a living room, refrigerator, laundry transfer, dryer start, and clothes folding, autonomously completing approximately 70 to 80 actions.

That kind of run is evidence of capability, not yet proof of a repeatable learning system. Sun Peng's proposed direction, reported by another outlet, is to connect reinforcement learning with real-robot training and deployment, so repeated execution and feedback improve task performance. Tencent is pursuing the algorithmic side too: woshipm reported that Pang Tianyu joined in January 2026 to lead exploration of frontier multimodal reinforcement-learning algorithms.

The shared challenge is turning real work, human intervention, simulation, and feedback into a disciplined data loop that makes each deployment more reliable.

A long demo is not a learning system

A long household demonstration can show that a robot can connect many actions into one visible sequence. According to geekpark, MORPHI KINO spent 15 minutes tidying a living room, checking a refrigerator, moving clothes from a dryer into a new load, starting the dryer, and folding clothes. The KINO robot autonomously completed approximately 70 to 80 actions. That is meaningful evidence of task coverage, but it does not by itself show a system that learns from what happens on later executions.

A demo records an outcome. A learning system needs a way to turn each execution into feedback that changes future performance.

Sun Peng's stated aim is to put reinforcement learning into real-robot training and deployment, allowing robots to improve task performance through repeated real-world execution and feedback, according to a report. The distinction matters because deployment is then part of the training process rather than the point at which training stops. A robot must act in the world, receive signals about the result, and use those signals in subsequent attempts. The polished sequence is only one moment in that loop.

The same direction appears in algorithm work. In January 2026, Tsinghua University computer science PhD Pang Tianyu joined Tencent to lead exploration of frontier algorithms for multimodal reinforcement learning, woshipm reports. Multimodal reinforcement learning points toward systems that can connect what a robot observes and does with feedback from execution. The practical question is therefore larger than reproducing a household routine: can each deployed run supply useful feedback, and can that feedback produce better task performance through repetition?

Build the data recipe around the task

A task should determine the data source, rather than a team treating teleoperation as a universal answer. IO-AI Tech separates the work into three product lines: TeleXperience for teleoperation, SenseXperience for real-world human data collection, and EmbodiFlow for data management. That division maps to distinct questions. A robot that must reproduce a controlled manipulation may need human-operated trajectories, while a task shaped by ordinary human motion can require observations of how people act in the setting.

The recipe changes with the scene.

SenseXperience captures real-world human operation through first-person-view units, wrist cameras and UMI grippers, according to the report. Those views can preserve the operator's perspective and hand-level action without presuming that a robot should receive identical demonstrations. TeleXperience, meanwhile, has been adapted to more than 80 robot models, including humanoids, robotic arms and mobile robots. That breadth matters because the useful record is tied to a robot's body, controls and working environment.

Deployment logs add another source: evidence of what the machine actually encountered after leaving a controlled collection session. But raw records are not automatically training material. EmbodiFlow covers collection, annotation, review, quality inspection, visualization and data-format conversion, the report says. The review stage is where a team can decide which attempted actions describe a reusable behavior and which merely document failure under unusual conditions.

Hardware choices also affect that judgment. According to geekpark, MORPHI develops its complete robot system and joint modules internally, while buying sensors and main control chips externally. It also relies on grippers for most current tasks rather than building dexterous hands in-house, because it sees algorithms as the main constraint. For such a system, the relevant data may center on perception, control decisions and gripper outcomes-not demonstrations designed around a hand it does not use.

Use real data to calibrate simulation

Real-Sim-Real treats physical collection and synthetic generation as linked stages, not competing sources of training material. Xingji Miyuan's loop begins with real-world collection, moves into simulation generation, then returns to real-world validation, according to a report. That final return matters because simulated scenes can extend a dataset without proving that a robot will behave correctly outside the simulator.

The quality of the simulated world depends on what was captured in the first place. A report says that Xingji Miyuan uses self-developed handheld collection devices and multimodal ego devices with lidar, cameras, thermal cameras, and 4D millimeter-wave radar. It also supports calibration across heterogeneous sensor combinations, with a calibration process controlled within 10 minutes. Field data therefore supplies more than examples for training: it gives simulation a measured reference for sensor relationships and scene reconstruction.

Validation closes the loop.

Tang Zheng of Cowin Capital argues that high-fidelity real-world reconstruction plus scalable simulation-generated training data is the most engineering-feasible route for embodied intelligence. Pan Yang, SwitchBot's CTO, similarly argues that the route broadens scenario coverage and enables fuller training and validation before deployment. The distinction is crucial: simulation can create variations that are difficult to collect, while real-world validation exposes where those variations produced misleading assumptions. Huang Qingqiu's account to geekpark of an end-to-end intelligent-driving system, spanning sensor input through planning and control trajectory between late 2024 and 2025, illustrates why the loop must preserve connections from perception to action rather than treating generated data as an isolated asset.

Embodied-AI data-and-feedback approaches across Chinese startups

Buy the plumbing, keep the learning judgment

The collection pipeline can be bought without outsourcing the decision about what counts as useful learning. According to a report, EmbodiFlow covers collection, annotation, review, quality inspection, visualization, and data-format conversion. It also supports LeRobot, MCAP, and HDF5.

That makes the plumbing more portable. A team can standardize how robot traces move from capture into review, then inspect data in a shared system rather than repeatedly building converters and labeling interfaces. But a clean file is not automatically a valuable example. The in-house question is which task state matters, which operator intervention reveals a failure, and which cases deserve another training round.

MORPHI offers a useful division of labor at the hardware level. geekpark reports that the company develops its complete robot system and joint modules internally, while buying sensors and main control chips externally. Its model is called MoRA, short for MORPHI Reasoning & Autonomy; MORPHI describes it as an Agentic-Native embodied foundation-model approach.

The same boundary applies to learning infrastructure. External tools can handle repeatable mechanics, while internal teams retain judgment over task definitions and failure interpretation. That judgment is closely tied to model work: woshipm notes that Tian Yonglong's OpenAI work centered on machine perception, generative models, and visually oriented representation learning. In January 2026, Pang Tianyu joined Tencent to lead exploration of frontier algorithms for multimodal reinforcement learning, according to woshipm.

Buying the pipeline does not buy the recipe. The durable advantage lies in deciding what the robot should learn from, what a bad execution means, and how reviewed evidence should change the next model.

Choosing a robot-improvement pipeline

• You need to collect demonstrations across a heterogeneous robot fleet rather than build a separate operator setup for every body. Consider IO-AI Tech's TeleXperience, SenseXperience and EmbodiFlow stack. TeleXperience has been adapted to more than 80 robot models, while SenseXperience collects real-world operation data through first-person-view units, wrist cameras and UMI grippers. Use EmbodiFlow when the workflow also needs collection, annotation, review, quality inspection, visualization and format conversion; it supports LeRobot, MCAP and HDF5. Do not treat the reported compatibility as proof that every robot will work without integration work.
• Your main gap is insufficient edge cases or scenario diversity before putting a robot back into the field. Use a Real-Sim-Real workflow: collect from the real environment, generate simulation data, then validate again in the real world. Xingji Miyuan explicitly follows this loop and is building around physical-asset reconstruction and high-fidelity scene simulation. This is the more suitable route when the target is training and validation across mines, industry or park inspection, rather than a narrowly bounded laboratory demonstration.
• A robot can complete an impressive scripted task, but the team lacks a mechanism for repeated post-deployment improvement. Prioritize a pipeline that captures execution and operator feedback, manages and inspects the resulting data, and feeds it into subsequent training. Stardust Intelligence's stated direction is to integrate reinforcement learning into real-robot training and deployment so task performance improves through repeated real-world execution and feedback. This is a strategy and stated technical direction, not a disclosed off-the-shelf product offering.
• You need a supplier for a specific industrial deployment rather than a general-purpose robot platform. Consider a project-based infrastructure provider. Xingji Miyuan expects to provide scenario customization, hardware-software integration, on-site deployment and subsequent operations and maintenance, as well as high-value scenario data through APIs or licensing. Do not choose this route solely because it claims one brain can support multiple robot types; that capability is a company statement rather than independently established performance evidence.

Measure reliability through recovery and validation

A reliability audit begins after the visible task sequence ends. According to geekpark, MORPHI KINO completed a 15-minute household task that ranged from tidying a living room to checking a refrigerator, then moving clothes from a dryer into a new load, starting the dryer, and folding clothes. geekpark counted approximately 70 to 80 autonomous actions. That is useful evidence of sustained execution, but action count alone cannot show what happened when an object was misplaced, a grasp failed, or the scene differed from expectation.

The audit should preserve those moments as training cases.

For each failed or uncertain execution, teams need a trace that links the task state to the attempted action, the recovery chosen, and the eventual result. The next model update should be checked against those captured cases, then returned to real-world validation rather than judged only in simulation. Xingji Miyuan's Real-Sim-Real loop-real-world collection, simulation generation, and real-world validation-offers the basic structure, according to a report. A rollback decision should follow if an update improves a narrow case while weakening recovery elsewhere.

Recovery is the harder reliability measure because it tests whether the robot can continue after its original plan breaks. Sun Peng said the aim is to bring reinforcement learning into real-robot training and deployment, allowing repeated real-world execution and feedback to improve task performance. That makes validation a continuing loop: run the task, capture the breakdown, update the policy, test on the robot, and retain or reverse the change. The practical record is not a polished completion. It is evidence that failures are detected, recoveries work, and later updates do not erase them.

Treat infrastructure commitments as a question about learning capacity, not a proxy for robot reliability. Alibaba's plan to invest at least 380 billion yuan in AI and cloud infrastructure over the next three years may expand compute access, but teams should ask how deployment data is retained, reviewed, and returned to training.

Also map who owns the model, product, and cloud layer. ByteDance separated Seed, Flow, and Volcano Engine around 2024, while Tencent moved some AI Lab staff into Hunyuan in March 2026. Those structures affect where failure reports go and who can change a policy.

Watch for evidence of sustained use rather than labels. Tencent said Hunyuan reached 131 products and internal Token usage rose about tenfold after a new model release. For embodied-AI vendors, request the equivalent trail: real-task records, quality-control decisions, model updates, validation results, and a defined recovery path when execution fails.

For readers outside China

• Availability: Availability outside China is not disclosed in sources. IO-AI Tech says its latest financing will support global market expansion, but the source material does not say where its products can currently be purchased or deployed. Xingji Miyuan describes project-based delivery and data access through APIs or licensing, but does not disclose international availability.
• Pricing: No product, API, licensing, deployment or subscription prices are disclosed in sources. Reported financing rounds are not product pricing.
• Closest Western equivalents: The source material does not identify specific Western equivalents.; LeRobot is the only named external interoperability reference: EmbodiFlow supports the LeRobot data format.
• Data residency: The source material does not cover data residency, data-export controls, customer ownership of collected demonstrations, cloud location, retention periods or cross-border transfer terms. Teams handling real-world human-operation data should obtain those terms directly before deployment.

Sources

• 36kr 前字节、腾讯AI核心研发孙鹏加入星尘智能，负责机器人强化学习后训练｜最前线 https://36kr.com/p/3966117507947780
• 36kr 36氪首发 | 腾讯、百度背景团队做具身数据基础设施，再获顺为等数亿元融资 https://36kr.com/p/3935994997407107
• geekpark 对话黄青虬：成立半年估值 70 亿，阿里腾讯联投，交出了能「长程执行」的第一个自研模型 https://geekpark.net/news/369724
• woshipm AI人事变阵：阿里收权，腾讯整军，字节留人 https://woshipm.com/it/6436080.html
• woshipm 腾讯混元多模态迎来新负责人，OpenAI田永龙将加盟 https://woshipm.com/share/6433365.html
• 36kr 36氪首发 | 南洋理工院士团队做全域具身智能系统，种子轮获同创伟业领投，卧安机器人、旭源资本跟投 https://36kr.com/p/3957129659497600