AUTOSIGNAL

Human-curated. Expert-annotated. Every signal traced to source. 人工精选 · 专家点评 · 每条信号可溯源

Research & IP研究与专利

Temporal-Distance JEPA: Learning Plan-Aware Latent World Models from Reward-Free Demonstrations时间距离JEPA：从无奖励演示日志学习规划感知的潜在世界模型

S 2.8 T1 1 sources1 个来源 R7-research cross-source×2

Temporal-Distance JEPA (TD-JEPA) targets a mismatch in Joint-Embedding Predictive Architectures (JEPAs): JEPA-style training optimizes only short-horizon latent prediction, while planning needs a multi-step ranking of imagined futures by goal progress — a ranking prior JEPA planners typically borrowed from latent Euclidean distance in embedding geometry.
Built on the existing LeWM encoder-predictor and SIGReg backbone, TD-JEPA mines a directed temporal cost directly from reward-free offline demonstration logs, using same-trajectory step order as positive targets, cross-trajectory pairs as heuristic negatives, and a rollout-consistency term to align the learned cost with the planner's horizon.
For autonomous driving, this offers a path to train latent model-predictive controllers directly from recorded driving demonstration logs, supporting trajectory planning and vehicle control without pixel-level reconstruction.

时间距离JEPA（TD-JEPA）针对联合嵌入预测架构（JEPA）的一个错配问题：JEPA式训练只优化短时程的潜在预测，而规划需要按目标进展对多步想象轨迹排序——此前的JEPA规划器通常直接借用潜在空间的欧氏距离作为排序依据。
TD-JEPA保留现有的LeWM编码器-预测器与SIGReg骨干网络，直接从无奖励的离线演示日志中挖掘有向时间成本：以同轨迹内的步骤顺序作为正样本目标、跨轨迹配对作为启发式负样本，并加入前向一致性项使学到的成本与规划时程对齐。
对自动驾驶而言，该方法为直接从记录的驾驶演示日志训练潜在模型预测控制器提供了路径，可支持轨迹规划与车辆控制，且无需像素级重建。

hf-daily-papers