Reproduce “Efficient Robotic Policy Learning via Latent Space Backward Planning” on OpenResearch

About this paper

Dongxiu Liu, Haoyi Niu, Zhihao Wang, Jinliang Zheng, Yinan Zheng, Zhonghong Ou, Jianming Hu, Jianxiong Li and 1 more

The paper introduces Latent Space Backward Planning (LBP), a framework that learns robotic policies for long-horizon, multi-stage tasks by predicting subgoals in reverse temporal order within a latent visual space. LBP achieves state-of-the-art success rates of up to 88.6% on the LIBERO-LONG benchmark and consistently outperforms forward planning methods in real-world scenarios by mitigating compounding errors.

Key insights

• Backward planning inherently mitigates compounding errors by anchoring all generated subgoals to the ultimate task objective, ensuring consistent 'on-task' alignment.
• Operating in a lightweight latent visual space for subgoal prediction significantly improves computational efficiency compared to pixel-level generative models.
• A unified, self-supervised recursive subgoal predictor enhances robustness against prediction errors during test-time inference by learning to predict from its own previous outputs.
• An adaptive goal-fusion module effectively processes diverse multi-horizon guidance from predicted subgoals and the final goal, allowing the policy to dynamically focus on relevant information.

Summary by alphaXiv.