Reproduce “SLA: Scaling Spatial-Language-Action Model via Continual Pre-training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou, Jingbo Wang, Junwu E, Pengwei Wang, Yan Wang and 2 more
SLA teaches a vision-language-action policy spatial structure by aligning its intermediate visual features with a frozen 3D perception model during continual pre-training and task-specific post-training, without adding depth as a policy input. On RoboTwin 2.0 Clean2Random, both-stage alignment reaches 72.3% success, above either single-stage alternative; this result supports the combined recipe in the tested setup, not universal causation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.