Humanoid Teleoperation / Imitation & Reinforcement Learning / MPC: How to Receive Everything from Deployment to Technical Support

💡 Here’s what you’ll learn in 5 minutes!

1. The Imitation/Reinforcement Learning + VR Teleoperation workflow that stole the spotlight at KRoC.

2. The step-by-step process: Teleoperation-based data collection → Imitation Learning → Reinforcement Learning.

3. How to run training and validation using Isaac Sim (Isaac Lab) tests.

4. How to secure comprehensive “Technical Support” through the Marosol Humanoid Special Exhibition.

Shifting Paradigms: The ‘Humanoid Learning Pipeline’ at KRoC

Bigwave Robotics (Marosol), a leader in robot automation, participated in the Korea Robotics Conference (KRoC) held at the Alpensia Convention Center in Pyeongchang, Gangwon-do, from February 4th to 7th.

At this exhibition, Bigwave Robotics showcased the entire R&D lifecycle essential for humanoid robot research: Remote Operation (Teleoperation) → Data Collection → Learning (Imitation/Reinforcement) → Simulation Verification. 

Rather than simply "displaying robots," we presented the actual learning infrastructure and operational know-how that can be immediately utilized in laboratories. This approach drew significant attention from academia and professional researchers alike.

Therefore, we have prepared today’s content in a Q&A format, featuring answers from Bigwave Robotics (Marosol) experts to the most frequent questions asked at KRoC by researchers, students, professors, and industry professionals considering the adoption of humanoid or research robots! 😎

 If you follow along to see “how humanoids learn, where they are verified, and how they transition to deployment,” the key points you've been curious about will become much clearer. 🙂

 The participation booth set up by Big Wave Robotics at KRoC (Korea Robotics Conference).

Q1. Hello, Doyeon. What did researchers and officials say they need most when introducing humanoids at this year's KRoC?

The overwhelming question on-site was: “How far does the technical support actually go?” In particular, researchers, professors, students, and practitioners at KRoC shared a common realization: a humanoid is not a piece of equipment where the journey ends with purchase and installation. 

To truly improve performance and achieve stable operation, a continuous 'R&D pipeline'—including remote control (teleoperation) setup, data collection, imitation/reinforcement learning, and simulation-based verification via tools like Isaac Sim—must keep running long after the robot is deployed.

 The Big Wave Robotics development team setting up a teleoperation environment at the KRoC expo.

 Q2. What Exactly Are Imitation Learning, Reinforcement Learning, and MPC?

That is a good question. In fact, when many people hear the term "technical support," they simply think of repairing a robot when it breaks down. However, the technical support referred to in humanoid robots is a much broader concept than that.

Unlike industrial robots that can be set up through simple teaching, humanoids are robots that can only move reliably in real-world environments when learning and control technologies are integrated.

Simply put, technical support for humanoid robots refers to  the entire process, from creating the brain that allows the robot to move to ensuring stable control so it doesn't fall over.

 It is helpful to understand this by dividing it into three main categories.

 1) Imitation Learning (IL)

Imitation learning is literally “a method of learning where a robot follows a human demonstration.”

For example, when a person directly operates a humanoid using VR Teleoperation, the robot accumulates data such as camera views (visual information), joint movements (angles, velocity), hand motions and force control, and task procedures (action sequences) during the process.

The data collected this way essentially becomes a “set of correct answers” from the robot's perspective, and by learning from it, the robot becomes capable of generating similar actions on its own in similar situations.

In other words, you can think of Imitation Learning as the technology that creates the starting point (baseline) for a humanoid to begin performing tasks from scratch.

 2) Reinforcement Learning (RL)

Reinforcement Learning (RL) is not a "method where humans provide the correct answer"; instead, it is a method where the robot improves itself by repeating trial and error.

For example, when performing a task like "grabbing an object," if you set up a reward system—such as +1 point for grabbing, -1 point for dropping, and +10 points for moving it to the target location—the robot gradually increases its success rate through thousands or tens of thousands of simulation iterations.

Because humanoids involve so many variables, it is difficult to rely solely on perfect teaching or imitation learning; only by creating a robust policy that is resilient to environmental changes through reinforcement learning can a robot be deployed in actual field environments.

Therefore, in practice, the most common approach is to quickly create basic movements with Imitation Learning (IL) and then boost performance and stability through Reinforcement Learning (RL).

 3) MPC (Model Predictive Control)

MPC (Model Predictive Control) is a slightly different domain from learning; it is the essential control technology required for a humanoid to move without falling over.

A humanoid is not just a robot that moves its arms; it must balance on its legs while simultaneously performing tasks with its hands. If the robot simply executes pre-learned movements, it might wobble or fall when faced with unexpected variables.

Simply put, MPC is a method where the robot predicts its movements for the next few seconds based on its current state and selects the most stable path among them.

In other words, it is a technology that makes the robot calculate the future while moving, rather than just reacting impulsively. For robots where balance is critical, like humanoids, MPC plays a vital role in enhancing the stability of actual physical movements.

 Q3. Then, to what extent does Marosol actually provide technical support?

A. The technical support we provide is not merely "troubleshooting when the robot stops moving." Instead, it includes support for designing and operating the entire 'post-deployment pipeline' so that the robot can actually learn and be validated in both labs and real-world fields.

In humanoid projects, the speed at which you build the learning environment and operational structure—rather than the robot itself—determines the success of the deployment. Therefore, Marosol provides practical support through the next stages of the humanoid implementation process.

 1) Support for establishing a tele-operation (VR remote control) environment

The first bottleneck in humanoid research is often “how to collect data.” Therefore, we provide a structure that allows you to intuitively control the humanoid through a VR-based teleoperation environment and naturally accumulate training data in the process. 

Simply put, we support researchers so they can immediately begin creating demonstration data and conducting experiments without having to develop a complex remote operation environment from scratch.

We ensure that the data collected during this process is not just simple video, but is organized into a “format ready for immediate training,” including visual information (camera/depth, etc.), joint movements (full-body joint trajectory), hand/gripper movements, and action sequences (task workflows).

 2) Support for Imitation Learning (IL) and Reinforcement Learning (RL) Training Pipelines

The data captured via VR Teleoperation is immediately utilized for Imitation Learning (IL), establishing the baseline that allows the robot to rapidly acquire fundamental movements from the start. We configure the learning environment to be compatible with open-source frameworks such as LeRobot and ROS2, supporting an open pipeline that isn't locked into a specific vendor or robot environment. This allows research institutions to naturally expand their humanoid studies without discarding their existing data, code, or experimental assets.

Building upon the basic movements created through IL, we then utilize Reinforcement Learning (RL) to boost the performance and stability required for actual field deployment. Since humanoids face many real-world variables, simply "mimicking" may lead to inconsistent success rates. Therefore, during the RL stage, we support the optimization of empirical factors—such as grasping robustness despite shifting object positions, balance maintenance, improved success rates for repetitive tasks, and adaptability to environmental changes—to align with your specific research goals.

 Visitors experiencing teleoperation with the Rainbow Robotics Humanoid RBY-1 at the KRoC expo

 4) Support for simulation verification based on Isaac Sim / Isaac Lab / MuJoCo

And one of the most crucial parts is simulation verification. Iterating learning directly on a physical humanoid is expensive and carries high risks, such as falls or collisions. 

Therefore, we support motion generation and pre-validation, including reinforcement learning, based on simulation environments like Isaac Sim, Isaac Lab, and MuJoCo. 

Through this process, research institutions can gain advantages such as ensuring experimental stability, improving repeat-test efficiency, and increasing the success rate of physical robot application. Ultimately, this completely transforms the speed of research itself.

 5) Integrated Support Extending from Deployment to Operation (EasyCare)

Lastly, humanoids are truly meaningful only when they move beyond the lab and connect to actual operational and verification stages.

That is why we don't just provide simple technical support. Instead, we offer EasyCare, an integrated support system that combines robot deployment consulting, research environment establishment, empirical operation support, technical maintenance, and the sharing of equipment operation expertise into a single flow.

Ultimately, for research institutions, this significantly reduces the "burden of solving everything alone after purchasing a robot," allowing them to focus entirely on the research itself.

In summary, Marosol's technical support is about “Success from Deployment and Beyond.” Buying a humanoid has now become less about "purchasing hardware" and more about integrating a complete learning pipeline.

This is why we at Marosol (Bigwave Robotics) have established a structure that practically supports the entire workflow: VR Teleoperation → Data Collection → Imitation Learning (IL) → Reinforcement Learning (RL) → Simulation Verification → MPC-based Stabilization → Real-world Application & Operation.

For this reason, at the KRoC (Korea Robotics Conference) site, the most frequent question was not about specific robot models, but rather: “How far does the technical support actually go?”

Q5. Why Do You Say "Operation" is More Important than "Purchase" for Humanoids?

In reality, “operating” a humanoid robot is much more difficult than “purchasing” one. Even after completing training with Imitation Learning (IL) and Reinforcement Learning (RL), and finishing validation in simulations, one final question remains when the robot enters the actual field: “Can our organization operate this robot consistently and stably?”

This is why Bigwave Robotics (Marosol) operates an Innovation Lab for humanoid demonstration. We aren't just focused on boosting learning performance; we are building an entire demonstration structure that connects directly to actual industrial site operations.

Furthermore, we believe the core of humanoid operation lies in Integrated Control and Orchestration. To support this, we are continuously advancing SOLlink, our cloud-based integrated control system that allows for the unified operation of diverse robot types on a single screen.

In the field, there eventually comes a moment where “the research team handles the learning, while the site handles the operation.” At that point, if there is no structure to connect the trained robot so it can run naturally within the actual work system, the success rate of the deployment will inevitably plummet. This is why designing everything from humanoid introduction to operational control is becoming an increasingly critical standard.

 Kim Min-kyo, CEO of Big Wave Robotics, standing with major research robots, including Humanoid robots, at the Innovation Lab.

The Marosol Humanoid Special Exhibition: Bridging Tech to Deployment

Once you understand the technology, the next step is simple: Shortening the timeline to find the right humanoid deployment route for your organization.

The Marosol Humanoid Special Exhibition is not just an informational showcase. It is a curated program designed to help you compare deployment-ready humanoid lineups → shorten delivery schedules → and receive integrated support for demonstration and operation—all in one consultation.

✅ Key Exhibition Benefits (Summary)

• Korea’s Largest Lineup: Compare 10+ types of deployment-ready humanoids from global and domestic brands.

• Average Deployment in 3 Months: Pre-orders can reduce wait times by more than half.

• PhD-level Research Support Team: Integrated support for everything from initial design and verification to operations.

• Customized Packages: Specialized deployment packages for research, industrial, and educational institutions.

👇 If any of these questions have crossed your mind, now is the time to receive your fastest deployment roadmap.

• “Can I get a comparative quote for robots suitable for my research?”

• “Can we start accumulating data using VR Teleoperation right away?”

• “Can we verify our specific tasks in Isaac Sim before the physical robot arrives?”

If you answered "Yes" to any of these, the fastest and most reliable way is to request a 'Deployment Roadmap' based on your target tasks, schedule, budget, and scope through our exhibition consultation. Click the button below to leave a brief inquiry, and a Bigwave Robotics expert will get back to you shortly. Thank you! 👇