Exact equivariance, kept through training, buys zero-shot generalisation across the symmetry group
精确等变性在训练中保持,实现跨对称群的零样本泛化
机构 * Department of Mathematics, Stony Brook University(石溪大学数学系)
专题命中 具身推理 :world model(abstract);分类 cs.RO、cs.AI、cs.LG
AI总结 通过等变编码器和预测器构建的潜世界模型,其训练损失具有可证明的对称性,从而在仅拟合部分方向动力学时,数学上确定整个轨道上的行为,实现跨对称群的零样本泛化。
Comments 112 pages, 19 figures. v2 adds programme lineage to companion papers (arXiv:2606.13092, 2606.24945, 2606.24946), engages the equivariance-at-scale debate (arXiv:2410.23179), and adds experimental hardening: 5-seed CIs, frame-averaging/canonicalization baselines, a real-robot DROID anchor, a scale-vs-exactness curve. Core claims unchanged. Code: https://github.com/TimothyWang418/se3-ejepa