Robots can be shown opening doors and dishwashers, but those demonstrations often depend on people steering the machine or walking it through the task. Researchers at ETH Zurich are exploring a way to reduce that hands-on work: give the system a high-level description, and let it determine how the robot should act.
From a description to a workable route
The process begins with a person describing the scene and the action the robot should perform. That instruction gives the system an overall goal, rather than a detailed sequence of movements.
The planner then develops a route, which may initially be more complicated than necessary. It refines that route into a minimal viable path: a shorter sequence intended to accomplish the task. This approach aims to limit manual guidance while leaving the robot to work out the practical details.
Planning movement, contact and force
The researchers describe the planner as using high-level information about the robot and the object, along with a task specification encoded through a sparse objective. From that input, it determines several connected parts of the task.
Those decisions include how the robot moves, which limbs it uses and what forces it applies. The system also plans when and where the robot should make contact with an object, and when that contact should end. These details matter because manipulating a door requires more than simply moving toward it: the robot must coordinate its body and its interaction with the object.
Two ways to frame a task
The system groups tasks into object-centric and robot-centric categories. Object-centric tasks focus on changing an object’s state, such as opening a door or dishwasher. Robot-centric tasks involve moving the robot itself around objects.
Separating these task types gives the planner a way to represent different goals. In one case, success centers on the object; in the other, it centers on the robot’s route through its surroundings. Both involve planning motion, but they ask the system to focus on different outcomes.
Demonstrations with a quadruped
The researchers say the system can be adapted to different robot designs. For their demonstrations, they used a quadruped, specifically ANYbotics’ ANYmal. The article notes that ANYbotics was spun out of ETH Zurich, making the robot a familiar platform for this kind of research.
The demonstrations show how the planning method can be applied to physical tasks, while the researchers’ broader aim is to use this work as a step toward a fully autonomous loco-manipulation pipeline. That means combining locomotion with the ability to manipulate objects, so a robot can move through an environment and interact with it as part of the same broader capability.
The work does not mean that robots already handle these tasks without human involvement. It points toward a process that can reduce the amount of guidance needed, by giving the system a goal and letting it plan movement, contact and force. If developed further, that could bring robots closer to opening doors without a person directing each step.