SLA scales spatial-language-action models via continual pre-training with frozen 3D alignment
HKUSTGZ · hf · 2026-11-06
HKUST(GZ), BAAI, Tsinghua and others present SLA, a Spatial-Language-Action model that scales implicit spatial alignment via continual pre-training. Initialized from a pretrained π0.5 policy, it aligns intermediate visual embeddings with spatial representations from a frozen 3D perception model, uses feature-dependent gates to aggregate multi-layer VLA features, exploits partially available depth and camera calibration with RGB, and applies probabilistic state dropout to encourage visual reliance. All auxiliary components are training-only, so the inference architecture is unchanged; heterogeneous robot datasets are unified for diverse embodiments, tasks, and scenes.
More from Embodied
- University of Michigan's 26-lecture robotics math course (ROB 501) is free on YouTube — techNmak · 2026-11-06
- Musk: Tesla to wind down Cybertruck production mid-next year to boost Cybercab output — elonmusk · 2026-11-06
- Xinhua hires two UBTECH UWORLD U1 humanoid robots as news anchors — CyberRobooo · 2026-11-06
- RoboColosseum benchmarks robot policies on real hardware across 20,000+ blind runs — DJiafei · 2026-11-06
- KAIST's DreamFree distills video diffusion representations for real-time robot control — kaist-ai · 2026-11-06
- Wayve details 'Learned Safety' approach to validate end-to-end AVs at scale — Jamie_Shotton · 2026-11-06