Unified Task and Motion Planning

This direction is split into other directions: task and motion planning integration. Its last version stays readable and citable here.

Version 1 of 162 papers67 conceptsfrozen 2026-10-04

In robotics, there is a fundamental tension between "thinking" and "doing." High-level reasoning allows a robot to sequence complex tasks, such as "make a cup of tea," while low-level geometric planning ensures the robot’s arm doesn't collide with a table. Traditionally, these two worlds were decoupled: a symbolic planner would generate a logical sequence of actions, and a motion planner would attempt to execute them.

However, as Arnav Jhala and R. Michael Young noted in Representational Requirements for a Plan Based Approach to Automated Camera Control, a massive gap exists between high-level logical planning and low-level geometric resolution. A plan can be perfectly logical—"pick up the tomato, then put it in the fridge"—yet physically impossible if a chair is blocking the fridge door or if the tomato is just out of reach. This research direction, known as Task and Motion Planning (TAMP), seeks to bridge this gap by unifying symbolic logic with continuous physical constraints.

The Logic Geometry Gap

The core challenge of TAMP is that symbolic planners often operate on abstract predicates (like In(Tomato, Fridge)) that fail to capture the nuanced physical realities of a robot's workspace. Caelan Reed Garrett explains in Sequence-Based Plan Feasibility Prediction for Efficient Task and Motion Planning that this disconnect leads to "computational waste." A robot might spend minutes computing a complex 10-step logical plan, only to realize at the very last step that a tiny obstacle makes the entire sequence unexecutable.

To solve this, researchers like Andre Gaschler and Ronald P. A. Petrick developed frameworks like Knowledge of Volumes (KVP) in Robot task planning with contingencies for run-time sensing. By representing robots and objects as sets of convex polyhedra, they allow the robot to "reason" about physical volumes and collisions during the high-level planning phase, rather than treating geometry as an afterthought.

Logic Geometric Programming

One of the most influential approaches to this unification is Logic-Geometric Programming (LGP), introduced by Marc Toussaint in Logic-Geometric Programming: An Optimization-Based Approach to Combined Task and Motion Planning.

Instead of treating task planning and motion planning as separate stages, LGP treats them as a single, unified optimization problem. The high-level logic provides the "structure" of the constraints (e.g., "the hand must be in contact with the mug"), while the low-level geometric solver finds the optimal continuous trajectory that satisfies those constraints. This allows robots to solve tasks where the goal is defined by a continuous objective—like "build the tallest stable tower possible"—rather than just reaching a simple symbolic state.

Integrating Samplers and Predicates

Another breakthrough came from Caelan Reed Garrett with PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning.

The intuition here is to use "blackbox samplers" to bridge the discrete and continuous. A symbolic planner might know it needs a GraspPose, but it doesn't know what that pose is. PDDLStream allows the planner to "call" a geometric sampler that generates potential physical coordinates. If a sampler finds a valid pose, it "streams" that information back into the symbolic logic. This "optimistic" approach assumes a solution exists and refines the plan as more geometric data becomes available, effectively interleaving logical search with physical verification.

Learning to Predict Feasibility

Even with integrated frameworks, TAMP is notoriously slow because checking physical feasibility (like inverse kinematics or collision checking) is computationally expensive. Recent work has turned to machine learning to accelerate this process.

In Learning Geometric Reasoning Networks For Robot Task And Motion Planning, Smail Ait Bouhsain and Rachid Alami introduced Geometric Reasoning Networks (GRN). These networks use Graph Neural Networks to predict whether an action is feasible in milliseconds. Crucially, they provide "interpretable hints"—explaining why an action might fail (e.g., "the object is blocked").

Similarly, Georgia Chalvatzaki and Tianyu Ren developed the Neural Feasibility Classifier (NFC) in Accelerating Integrated Task and Motion Planning with Neural Feasibility Checking. By filtering out geometrically impossible plans before the robot even attempts to move, these neural heuristics can reduce planning time by over 60%.

Why It Works

The success of Unified Task and Motion Planning stems from its ability to treat the robot's world as a single, coherent entity rather than a collection of disconnected abstractions. By:

  1. Informing logic with geometry early (as seen in RHH-LGP),
  2. Abstracting constraints into symbols (as seen in U-TAMP), and
  3. Predicting failures through experience (as seen in PIGINet),

TAMP allows robots to move beyond simple reactive behaviors toward true, long-horizon autonomy in complex, cluttered environments.

Papers in this version

Showing 60 of 62 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_unified_task_and_motion_planning_75597c0b,
  title = {Unified Task and Motion Planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unified-task-and-motion-planning/rdr_75597c0b68414fd58c1a1441305766a6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_unified_task_and_motion_planning_75597c0b_v1,
  title = {Unified Task and Motion Planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unified-task-and-motion-planning/rdr_75597c0b68414fd58c1a1441305766a6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}