[KongchangAI]
Model

AtomTree

A hierarchical, state-aware framework for LLM-based embodied planning that achieves robust planning through the incremental construction of a tree of subgoals and actions.

Timeline (last 90 days)

Oct 30

AtomTree is a hierarchical framework for LLM-driven embodied planning that incrementally builds a tree of subgoals and actions and makes decisions based on the current symbolic state.

Unverified50%
Oct 30

AtomTree introduces a lightweight memory module that summarizes the execution results of past subgoals and reuses them in subsequent decisions, without storing raw trajectories.

Unverified50%
Oct 30

When low-level action execution stalls, AtomTree triggers hierarchical backtracking replanning: it returns to the parent node and regenerates the subtree plan based on the latest state.

Unverified50%
Oct 30

In zero-shot evaluation on the ALFRED benchmark, AtomTree outperforms prior LLM-based planners.

Unverified50%
Oct 30

AtomTree's incremental decision-making strikes a compromise between open-loop and closed-loop planning: it retains the LLM's reasoning ability while introducing closed-loop feedback through state awareness and backtracking replanning, balancing robustness and efficiency.

Unverified50%

All Facts (5)

Source Articles