AI news ·
SLA Robot Policy Beats Pi 0.5 by 24 Points in Real-World Tests
A new robot training method combining action prediction with 3D perception achieved an 83.8% real-world success rate, outperforming a prior policy by 24 points.
A new robot policy called SLA achieved an 83.8 percent average real-world success rate in tests, outperforming the Pi 0.5 policy by 24 percentage points. The method combines action prediction with frozen 3D-perception features during a continual pretraining phase.
The training corpus spans 1.7 million episodes across 1,504 distinct tasks. The results indicate that enforcing spatial alignment between perception and action can materially improve robotic reliability in physical environments.