Research Analyzer
← Back ICRA 2026

Anticipatory Task and Motion Planning: Improved Rearrangement in Persistent Continuous-Space Environments

Roshan Dhakal, Duc Nguyen, Tom Silver, Xuesu Xiao, Gregory Stein

PDF

AI summary

Key figure (auto-extracted from paper)
Anticipating future tasks during planning significantly reduces overall cost and improves performance for long-lived robots in persistent environments.
Anticipatory planning Task and motion planning Graph neural networks Persistent environments Robot rearrangement Learning-augmented planning

Problem

Existing myopic task and motion planning strategies ignore future tasks, causing side effects that impede subsequent rearrangement tasks in persistent continuous-space environments.

Approach

A learned graph neural network estimates expected future costs, which are combined with an off-the-shelf TAMP planner to select plans that minimize both immediate and anticipated future costs.

Key results

  • 32.7% average per-task cost reduction in navigation domains
  • 16.7% average per-task cost reduction in cabinet-loading tasks
  • Up to 83.1% cost improvement with advance environment preparation
  • Successful real-world deployment on a Fetch mobile manipulator

Why it matters

Enables long-lived robots to operate more efficiently in dynamic, persistent environments by proactively avoiding planning side effects.

Abstract

We consider a sequential task and motion planning (TAMP) setting in which a robot is assigned continuous-space rearrangement-style tasks one-at-a-time in an environment that persists between each. Lacking advance knowledge of future tasks, existing (myopic) planning strategies unwittingly introduce side effects that impede completion of subsequent tasks: e.g., by block- ing future access or manipulation. We present anticipatory task and motion planning, in which estimates of expected future cost from a learned model inform selection of plans generated by a model-based TAMP planner so as to avoid such side effects, choosing configurations of the environment that both complete the task and reduce overall cost. Simulated many-task deploy- ments innavigation-among-movable-obstaclesandcabinet-loading domains yield improvements of 32.7% and 16.7% average per- task cost respectively. When given time in advance to prepare the environment, our learning-augmented planning approach yields improvements of 83.1% and 22.3%. Finally, we also demonstrate anticipatory TAMP on a real-world Fetch mobile manipulator.

Index terms

Integrated Planning and Learning Task and Motion Planning

Related papers