SLA Lifts Spatial-Language-Action Robot Policy 24pp Over Pi 0.5 Across Real-World Deployment

OpenHelix Team's SLA jointly optimizes action prediction and alignment against features from a frozen 3D perception model during continual pre-training on ~1.7M episodes across 6 datasets and 1,504 tasks. Across 8 real-world robot conditions SLA averages 83.8% success versus 59.8% for Pi 0.5, a 24pp lift, and reaches 68.4% on RoboCasa-Panda with only 50 demonstrations per task — above baselines trained with 10-60x more data.