Humanoid Robots Can Now Take Plain-English Orders to Open Doors
One spoken instruction, and a robot cranks a winch hard enough to lift another robot.
Humanoid robots have gotten good at walking, balancing and reaching for things. But the moment a robot actually touches something — a doorknob, a box, a crank — things get tricky. Push too gently and nothing happens. Push too hard and you break the handle, or the robot's arm. Until now, engineers had to hand-tune each robot for each task, which is slow and doesn't scale.
A team of researchers at Carnegie Mellon and elsewhere built a fix called KPI. Think of it as a translator sitting between the instruction and the robot's muscles. Instead of telling the robot exactly how to move, you give it a simple 'contract': for each direction, follow the path, give way, or hold a certain amount of force. The system reads how much resistance it's feeling — measured many times per second — and constantly adjusts how stiff or soft the robot's arms should be. It's a bit like how you instinctively loosen your grip when a jar lid starts to slip.
The clever part is that a vision-language AI (an AI that can see and read instructions) writes both the plan and the contract from a single sentence, with no task-specific programming. In demos, the robot opened doors, carried boxes, and operated a winch — turning a crank to hoist a second robot completely off the ground.
The catch: this is a research paper, not a product. The demos happened in controlled lab settings, and the robots still need to be set up carefully. Don't expect a humanoid helper in your home anytime soon. But the direction is clear — robots are moving from performing memorized motions to handling the messy, unpredictable physical world the way we do.
- KPI is a software layer that lets a humanoid robot take a plain instruction like 'open the door' and work out the physical force needed on its own.
- It adjusts the robot's grip and stiffness in real time, similar to how you loosen your hold when a lid starts to slip.
- In one demo, a robot turned a crank to lift another robot fully off the ground — no task-specific code required.
Why It Matters
Robots that can handle real-world contact from simple instructions could eventually do warehouse, factory and home chores now done by people.