Microduck makes robot training a user workflow

Microduck turns a robot from hardware into a training loop: a neural-network policy can move from MuJoCo simulation onto a physical machine. According to woshipm, Pollen Robotics supplies pre-trained motion policies and has opened the SDK, simulation environment, and reinforcement-learning tools that let developers alter that loop.

The same pattern appears in coaching products. One report says that Huanke APP records error rates and exit points to produce child-specific learning reports, while Heygo uses snowboard and foot sensors to drive voice feedback during skiing. The useful comparison is not which device seems smartest. It is how behavior becomes data, how that data triggers the next action, and who can inspect or change the resulting system.

From sensor reading to the next action

Microduck's loop begins before the robot moves. Pollen Robotics supplies pre-trained motion policies and has opened its SDK, simulation environment, and reinforcement-learning tools, according to woshipm. Developers can train in MuJoCo, then export a neural-network policy for deployment to a physical Microduck. The useful output is not a sensor reading on its own; it is a policy that determines the robot's next movement.

Heygo uses the same basic conversion in a sport setting.

For model training, Heygo adds sensors to snowboards and captures movement from both feet as well as the board itself. That data lets its model infer snowboard movement from foot data. The system returns its interpretation while the user is skiing through Bluetooth-earphone voice feedback, then provides a fuller review during lift rides. At the end of the day, the user receives a personalized report with training recommendations.

Huanke APP turns learning behavior into a different kind of output. It records time spent on each animation frame, repeated errors on each question, exit points, and personalized feedback, then generates a multidimensional learning report for each child. Its AI spoken-English tutor, incorrect-question collection, and phased learning suggestions make the report actionable. In each case, capture is only the first step: the product earns its role by structuring observed behavior and sending back a next action.

When measurement earns the right to interrupt

Microduck makes repeated experimentation feel unusually accessible: the robot is 25 centimeters tall, weighs less than 800 grams, and costs $399, according to woshipm. Its demonstrated actions include walking, kicking a ball, roller skating, retrieving objects with its bill, and recovering after a fall. The official training example can operate across 4096 simulation environments at once. With a CUDA GPU, woshipm reports an estimate of about 1-2 hours for a usable walking gait.

That speed does not, by itself, earn a device the right to correct a person in motion.

Heygo's ski-coaching system shows the stricter standard for an interruption. Across two ski seasons in the Northern and Southern Hemispheres, the company tested dozens of sensors at resorts in temperatures down to minus 30 degrees Celsius and gathered more than 1 million records of real skiing data. Sixty percent of its team has international ski-instruction certifications. Heygo says its current data accuracy is within 3 degrees.

The task and timing are also explicit. Heygo sends real-time voice feedback through Bluetooth earphones while a skier is skiing, saves a fuller review for lift rides, then provides an end-of-day report with training recommendations. That division matters: a correction during movement must be accurate enough to trust, while a more detailed explanation can wait for a safer pause. Microduck's rapid simulation loop is useful for generating robot behavior; Heygo's field data illustrates what is needed before a system interrupts human behavior.

How three AI devices turn captured behavior into feedback or model training

Open training and proprietary behavioral moats

Microduck's training proposition begins with materials that can circulate beyond the robot itself. Pollen Robotics supplies pre-trained motion policies and has opened Microduck's SDK, simulation environment, and reinforcement-learning tools, according to woshipm. Developers can train in the MuJoCo physics simulation environment, then export a neural-network policy for deployment on a physical Microduck. LeRobot, an open-source robot-learning toolkit, lets developers share models, datasets, and training methods. As geekpark's Pan Hao argues, the hardware can therefore act as an embodied data platform for models trained toward different purposes.

The advantage is modifiability. A developer can change work in simulation and circulate the resulting training artifacts through LeRobot.

Heygo pursues a different advantage. Its proprietary data-labeling platform covers the path from collection through validation, reports say. The company claims more than 1 million real skiing data records, plus a snowboard knowledge graph containing 15 to 20 advanced metrics. Fengluan Technology intends to use collected data to create a longitudinal growth database for each child spanning 6 to 12 years. These systems may build depth that an open toolkit does not own, but that value is concentrated in company-held records and profiles rather than in shareable policies or datasets.

The conflict is not simply open versus closed. Microduck makes training methods and deployable policies more available for modification; Heygo and Fengluan seek an advantage from behavioral data that remains inside their systems. The practical distinction is where learning can be edited, and where it becomes proprietary memory.

A child profile is more sensitive than a robot policy

Fengluan Technology's desktop learning companion robot makes the child's body and reactions part of the learning record. According to reporting, it relies primarily on voice interaction, while a camera's built-in model monitors sitting posture, micro-expressions, concentration, and emotional feedback. Parents and teachers can use mini-programs to synchronize learning progress and reports in real time. The proposed system therefore reaches beyond answers or lesson completion.

It builds a profile from conduct as well as coursework.

Fengluan's planned hardware expansion raises the stakes. Reporting describes plans for children's health wristbands and sports hardware that would collect information about learning, behavior, health, and social interaction. The company also plans a longitudinal growth database for each child spanning many years. That duration changes the question from whether a device can offer useful coaching to what should happen when a child's earlier posture, attention, or emotional signals remain available long after the immediate learning moment.

Meaningful consent needs to address the profile, not merely the device. Parents and teachers may need clear control over what is collected, which records can be corrected, and when a child's data can be deleted. Minimization matters because a report shared in real time can be useful without every underlying signal becoming a long-lived asset. Limits on secondary use matter for the same reason.

A connected learning system earns trust when its collection boundary is understandable and its exit path is real.

Choose the device by the quality of its feedback loop

• You are a robotics developer who needs to create, inspect, and replace a physical robot's behavior rather than merely use a preset feature. Choose Microduck. Pollen Robotics provides pre-trained motion policies and has opened its SDK, simulation environment, and reinforcement-learning tools. Developers can train in MuJoCo and deploy an exported neural-network policy to a physical Microduck; models, datasets, and training methods can also be shared through LeRobot. Do not choose it solely as a finished consumer appliance if you do not expect to use that training workflow.
• You have a CUDA-compatible NVIDIA GPU and want a relatively fast simulation-to-hardware experiment. Microduck is the clearest fit in the sources: its official example can run 4096 simulation environments simultaneously, and the official estimate for a usable walking gait with a CUDA GPU is about 1-2 hours. Without a suitable GPU, training can be submitted to cloud GPUs through Hugging Face Jobs, but cloud pricing is not disclosed in sources.
• You want ski coaching during a session and a structured review afterward, and are willing to use boot-mounted sensors and Bluetooth earphones. Consider Heygo. Its mass-produced product uses two sensors on the outside of ski boots, gives real-time voice feedback through Bluetooth earphones, provides a complete review during lift rides, and produces a personalized report with training recommendations at the end of the day. Its claimed accuracy is within 3 degrees. Do not treat that figure as independently verified, because it is Heygo's claim.
• You are assessing a learning device for a child and want progress reports across home, parent, and teacher contexts. Huanke APP is the more established part of Fengluan Technology's offering: it records time spent on each animation frame, repeated error rates for each question, exit points, and personalized feedback to generate multidimensional learning reports. Parent and teacher mini-programs synchronize progress and reports in real time. Do not buy the desktop robot for immediate use based on the source material alone: the company expects to complete its first-generation prototype before November this year.
• You want to minimize collection of sensitive behavioral data, especially video-derived attention, expression, or emotional signals. Avoid assuming these systems are data-minimizing. Fengluan Technology's planned desktop learning companion robot uses a camera model to monitor sitting posture, micro-expressions, concentration, and emotional feedback, and the company plans a longitudinal growth database spanning 6 to 12 years. Heygo's product is explicitly built around motion-data collection and analysis. The sources do not say what user controls, deletion options, or retention limits are available.

Check the exit path before buying the coaching

A coaching device should be evaluated at the point where its service stops being convenient. Heygo illustrates the dependency: its hardware includes a basic version, but its in-depth analysis and personalized plans require additional payment, according to a report. Its loop also depends on timed access to feedback: voice guidance arrives through Bluetooth earphones while skiing, a review comes during lift rides, and an end-of-day report supplies recommendations.

Ask what remains if the subscription ends.

That question applies differently to Microduck. According to woshipm, developers with a CUDA-compatible NVIDIA GPU can train it locally; those without suitable hardware can send training work to cloud GPUs through Hugging Face Jobs. Local training offers a clearer continuity path, but only if the collected demonstrations, model files, and training setup can be retained and understood outside a hosted service. Before committing, ask for the export format, the location of raw data, and the time needed to correct a bad example before it affects the next training run.

Open tooling changes the bargaining position. LeRobot is an open-source robot-learning toolkit through which developers can share datasets, models, and training methods, woshipm reports. That does not guarantee every project is portable, yet it gives buyers a concrete alternative to a single vendor's interface. A useful test is simple: can the same data and policy move into a workflow the buyer controls?

Ownership and access terms can shift around cloud infrastructure. woshipm notes that The Information reported an agreed NVIDIA acquisition of Hugging Face for $12.9 billion, while Business Insider reported that no final agreement had been signed and the transaction could change. NVIDIA was already an investor in Hugging Face's 2023 financing round. The reports do not establish an outcome. They do show why continuity planning should precede dependence: identify the cloud step, the paid feature, and the export path before the coaching becomes part of daily practice.

Before adopting a coaching device, test the gap between what it measures and what you can correct. Heygo says its accuracy is within 3 degrees; users should check whether that precision produces advice they can act on while skiing, rather than only a polished report afterward.

Its feedback arrives through Bluetooth earphones during a run, expands into a lift-ride review, and ends with training recommendations. That makes the timing of correction part of the product. Also inspect the payment boundary: Heygo includes a basic version with hardware, while in-depth analysis and personalized plans cost extra. Ask what data feeds those paid recommendations, how the system's labeling and validation process affects errors, and what happens to records collected across repeated sessions.

For readers outside China

• Availability: Microduck is manufactured in Shenzhen and sold directly to consumers, with few intermediaries, according to Seeed Studio's Pan Hao. Orders are scheduled as far ahead as next year. Heygo expects presales in September and shipment in October, with its first batch covering 38 countries and regions; it plans to sell through an independent website, domestic e-commerce channels, and more than 20 offline ski-equipment stores worldwide. For Fengluan Technology's Huanke APP and planned learning companion robot, availability outside China is not disclosed in sources.
• Pricing: Microduck costs $399; Reachy Mini also costs $399. Fengluan Technology's desktop learning companion robot is priced at around 6000 yuan. Heygo includes a basic software version with the hardware, while in-depth analysis and personalized plans require additional payment; the hardware price and subscription price are not disclosed in sources.
• Closest Western equivalents: Microduck most closely resembles an open-source robotics development platform: developers can train in MuJoCo, export a neural-network policy to the physical robot, and share models, datasets, and training methods through LeRobot.; Heygo resembles sensor-based sports analytics and coaching products, but its specific loop combines two boot-mounted sensors, real-time audio feedback, lift-ride review, and an end-of-day personalized report.; Huanke APP and Fengluan Technology's planned robot resemble an AI tutoring and learning-analytics system paired with a home learning device, using granular interaction records and parent-and-teacher progress synchronization.
• Data residency: The source material does not cover where Microduck, LeRobot, Huanke APP, Fengluan Technology, or Heygo store or process data, whether data leave the country where a user lives, or what deletion, export, parental-consent, or retention controls exist. This is especially material for Fengluan Technology's stated plan to collect learning, behavior, health, and social-interaction data for a longitudinal growth database spanning 6 to 12 years, and for its planned use of camera-derived posture, micro-expression, concentration, and emotional-feedback signals.

Sources

• 36kr 产品观察 | 前字节、大疆、腾讯团队创业，瞄准AI滑雪赛道，获厚雪投资 https://36kr.com/p/3965025284070664
• geekpark 专访爆火「机器鸭」背后的硬件推手：这是个信号，未来推动新故事的并非硬件 https://geekpark.net/news/370269
• woshipm Hugging Face"赶鸭子上架"，只要399美元 https://woshipm.com/it/6457839.html
• 36kr 半年融资三轮，这家公司以AI和数据切入儿童成长交互赛道丨36氪首发 https://36kr.com/p/3975552848589056