[No Filter Robot]Will Imitation Learning Break Robots' Speed Limits?

Director Park Hyunjun of Robeuroseu Wins Best Paper Award at the Korea Robotics General Academic Conference

Can robots that dance and run like humans evolve into robots that surpass humans? A research paper proposing a direction to answer this question has emerged in South Korea.

Park Hyunjun, director at Robeuroseu, is holding up the certificate after winning the Best Paper Award at the 2026 Korea Robotics Comprehensive Academic Conference on February 6. Photo by Paek Jongmin

Park Hyunjun, Director at Robeuroseu, received the Best Paper Award at the 2026 Korea Robotics Comprehensive Academic Conference held in Pyeongchang, Gangwon Province, on February 6, for his paper titled "Proleptic Temporal Ensemble for Improving the Speed of Robot Tasks Generated by Imitation Learning." Robeuroseu is a domestic robotics startup that develops humanoids, robot hands, and imitation learning-based object manipulation technologies.

Park explained that the starting point for this research was the "limit of speed." He stated, "Imitation learning enables robots to learn directly from data demonstrated by humans. The problem is that the robot's work speed ultimately becomes restricted by the demonstrator's speed."

Imitation learning, which enables robots to learn behaviors through human demonstrations, is drawing attention as a core technology for generating robot actions in unstructured environments. However, conventional methods are structurally dependent on the speed at which humans perform tasks. According to Park, aside from having demonstrators move faster during the data collection stage, there have been no clear alternatives for increasing robot speed.

To overcome this limitation, he proposed the "Proleptic Temporal Ensemble (PTE)" technique. Regarding this method, he explained, "The key is to predict future actions and bring those goals forward to the present for execution. Instead of simply following humans step by step, the robot moves toward a goal that is one step ahead."

PTE can be applied on top of existing Transformer-based Action Chunking (ACT) algorithms and directly utilizes already collected demonstration data and pre-trained policies. Its main features are that it does not require additional training or separate computational costs, and the implementation is relatively simple.

The performance was verified through real-world robot experiments involving block color classification. Compared to the ACT method, PTE maintained a high success rate while improving task execution speed by up to three times.

Regarding the impact of this research, Park noted that ▲ it can increase the productivity of existing policies without collecting additional data, thereby reducing the costs of industrial applications; ▲ by technically extending the performance ceiling of imitation learning-based robots that previously remained at human levels, it can broaden the scope of automation in logistics, manufacturing, and service sectors; and ▲ by providing a design framework that balances speed and stability, it lays the foundation for future research into high-speed, high-precision autonomous object manipulation technologies.

He concluded his acceptance remarks by stating, "The goal was to break the premise that imitation learning-based policies are bound by the demonstrator's speed."

This content was produced with the assistance of AI translation services.