Robot Training Breakthrough: 80% Less Human Coaching, Same Results

A team of robotics researchers has published a new method, PHIRL, that teaches robots new skills using far less human supervision. Today, getting a robot to fold laundry or stack boxes usually means recording many full demonstrations and labeling them carefully. That is slow, expensive, and it is a big reason robots are not in most homes yet. PHIRL cuts that labeling work by around 80 percent while actually improving how well the robot performs.

The trick is changing what humans are asked to provide. Instead of correcting every single movement, a person just marks progress — how far along the task is, like a loading bar on your screen. The system combines those progress marks with the demonstrations it already has, then repeatedly adjusts the robot's internal scoring rule (the 'reward' it chases) until it matches reality. In plain terms, the robot learns what a good job looks like rather than being told move by move.

The team tested PHIRL on real and simulated robots, and also used a vision-language model (an AI that can look at images and describe them) to generate progress feedback automatically. The method beat existing approaches on task success using only 20 percent of demonstrations annotated. Just as importantly, they tested 'reward hacking' — when a robot finds a shortcut that scores points without finishing the job. PHIRL's scoring rules held up better than the alternatives.

The catch: this is research, not a product. Tests happened in labs and simulations, and robots still needed human progress feedback, so a person stays in the loop. Real homes are messier than labs, with clutter, pets, and surprises. Still, expensive training is the bottleneck that keeps helpful robots costly, and this points toward machines that learn your routines from a few minutes of coaching instead of a full training camp.