Event

Guanchen Du et al. propose search-free planning framework HIJP

1 reports · 1 sources

Guanchen Du et al. proposed the Hindsight Inverse JEPA Planning (HIJP) framework on arXiv, which uses a goal-conditioned inverse model to directly predict actions from current and goal observations, replacing the costly test-time action sequence optimization in latent world models. At inference time, the inverse dynamics model and the latent prediction model are applied alternately to generate multi-step action sequences without online search. The paper claims that on four visual control tasks, HIJP achieves an overall success rate of 86.8%, and planning speed is 1.9–16.5 times faster than search-based planning; during training, future observations on offline trajectories are relabeled as goals, providing action supervision for multiple time spans.