A joint DGIST–KAIST research team has unveiled a training method for a small AI that jointly handles human-motion prediction and robot path planning.
To avoid pedestrians, a robot must read their preceding movements while simultaneously determining a safe path for itself.
The problem is that when these two decisions are incorporated into one model, their learning can interfere with each other.
3-line summary
1. Even a small AI can handle both decisions together
2. DPT reduces learning interference between tasks
3. Real-robot application remains a stage for future validation
Calculating a Person’s Next Movement and the Robot’s Path Together
Robots operating in crowded spaces need two decisions at the same time. They must predict which direction nearby people will move and, based on that prediction, establish a travel path that avoids collisions with them. Getting only one side right is not enough. Misreading a person’s movement can make the path dangerous, while inaccurate path planning can prevent the prediction from being connected to safe movement.
Assigning each task to a separate AI model requires more computing resources and memory. This is why it is difficult to load multiple large models onto an actual robot. A method is therefore needed to combine prediction and planning functions in one small model, but in this setup, each task may use the same parts inside the model and interfere with the other’s learning.
The DGIST research team defined this phenomenon as a skill collision. The term means that functions that must work together to achieve the external goal of collision avoidance can instead reduce each other’s performance inside the model.
Training Separately, Then Combining Only the Necessary Parts
The method proposed by the research team is Disentangled Parameter Training (DPT). Each function is trained primarily on different parts inside the AI, after which the parts required for each function are selected and combined into one model. The design aims to reduce interference between prediction and path planning without substantially increasing the model’s size.
In validation using JRDB and JTA, representative datasets in the field of robot movement, the research team said that the accuracy of predicting nearby people’s movements improved compared with existing methods, while robot path error and the likelihood of collisions decreased. According to a ZDNet Korea report, compared with DTPP, a representative integrated prediction-and-planning model, the average robot path error decreased by approximately 41.6%, and the collision rate decreased by approximately 49.0%. The average error in predicting human movement also reportedly decreased by approximately 33.7%.
These figures are the results of comparative experiments by the research team. Rather than interpreting them as guaranteeing the same level of reduction in every real-world environment, their significance lies in showing the potential to reduce interference when two tasks are handled together within a small model.
Confirming Potential Applications to Autonomous Driving
The research team also conducted additional experiments with autonomous-driving AI. After applying DPT and sparse merging, they reported that the driving-path completion rate and overall driving score in the Bench2Drive evaluation were higher than those of the existing HiP-AD model, while the average planning-path error decreased from 1.10m to 0.99m.
Robot News reported that the technology’s potential for use in the physical AI fields of robotics and autonomous driving had been confirmed. However, Professor Daehee Park of DGIST said that the plan is to apply it to actual robots in the future and verify its potential for use. Therefore, this achievement is not an announcement that field application has been completed; it is a stage at which a method for reducing the internal conflicts that arise when a small AI handles multiple decisions has been presented through experiments.
One point to note from this research is that simply giving AI more functions is not enough. For a robot to move safely alongside people, prediction and planning must not only work well individually; they must also be designed so that the two functions do not interfere with each other even inside a small device. The research was accepted to ECCV 2026, and the conference was held in Malmö, Sweden, from September 10 to 12.
References
Tags #RobotAI #CollisionAvoidance #SkillCollision #DGIST #KAIST #DisentangledParameterTraining #DPT #AutonomousDrivingAI #PhysicalAI #RobotPathPlanning #PedestrianPrediction #ECCV2026