---
kind: concept
slug: manipulation-taxonomy
title: Manipulation taxonomy
summary: >-
  How robot manipulation is classified and counted. T-DOM tags each step of a deformable-object task by deformation, motion energy regime and contact; DexGrip compares gripper designs by degrees of freedom. Tags are hand-made with no agreement score, and DoF counts follow each author's convention. Author-reported.
tags: [manipulation, taxonomy, deformable-objects, grippers, degrees-of-freedom]
papers: ["2411.17124", "2412.20998"]
related: [in-hand-manipulation, task-decomposition, robotic-grasp-generation]
status: draft
updated_at: 2026-10-25
---

## Overview

A manipulation taxonomy is a shared vocabulary for what a robot does to an
object and what its hand can do. Two cards use it in different ways. T-DOM
labels every step of a task with tags [arxiv:2412.20998 §4]. DexGrip places
its gripper in a table of degrees-of-freedom (DoF) counts for six other
grippers [arxiv:2411.17124 tab. I]. Neither card reports a downstream robot
benchmark, so the taxonomies here describe designs and demonstrations, not
controllers.

## Variants

### Deformation-aware action tags (T-DOM)

T-DOM targets objects that change shape, such as cloth, rope or foam. It has
three branches: deformation (D), robot motion (M), and interactions. The
interactions split into prehensile grasps, non-prehensile contacts and
contact sliding [arxiv:2412.20998 §4]. Deformation is tagged by force type
(compression, tension, torsion, shear) and by bending, which has a structured
level and an unstructured level [arxiv:2412.20998 §3.1] [§4.1]. Motion uses
the kinetic, elastic, gravitational and gravitational-elastic classes [§4.2].
Bimanual tasks carry a separate tag set per arm [§4.6]. Card:
[T-DOM](/papers/2412.20998).

On ten recorded tasks, the authors report 4 action clusters with more than
one member. Their own baselines, Bullock 2012 and Paulius 2020, give 21 and
22 clusters on the same actions, and the largest clusters are 10 and 17
[arxiv:2412.20998 §5.3]. These counts come from the authors' own labels, and
the paper reports no second annotator [arxiv:2412.20998 §5.2].

### Degrees-of-freedom count (DexGrip)

DexGrip is a soft gripper with three Fin Ray fingers, each with a belt, and
a telescoping palm with a rotating suction cup. Its table lists DoF counts
from 3 to 14 across six cited grippers, with DexGrip at 9 [arxiv:2411.17124
tab. I]. The 9 comes from three finger bends, three belt rotations and three
palm motions [arxiv:2411.17124 §III-A]. The palm count includes suction, which
is not a DoF in the usual sense, so the 9 is the authors' convention
[arxiv:2411.17124 §III-A]. Card: [DexGrip](/papers/2411.17124).

## When to use which

| Situation | Use | Why |
|---|---|---|
| You label demonstrations of a deformable object and need the labels to agree across annotators | T-DOM tag set, with your own agreement check added | The tags are defined, but the paper gives no labelling protocol and no agreement score [arxiv:2412.20998 §5.2] |
| You want to compare gripper designs by how many independent motions they have | DoF count, as in the DexGrip table | Counts for the other grippers are taken from their papers. Each author counts differently, so state your rule [arxiv:2411.17124 tab. I] [§III-A] |
| You are choosing a gripper that must re-orient an object without re-grasping | DexGrip as a design reference, not as a measured choice | Its in-hand tests are one prototype, judged by eye, with no trial counts [arxiv:2411.17124 §IV-C] |
| You need to compare task segmentations with existing taxonomies on the same actions | T-DOM's cluster rule, plus the baseline counts | The counts are cluster numbers on one labelled set, not a benchmark score [arxiv:2412.20998 §5.3] |

## Implementation sketch

The core of T-DOM's clustering is a grouping by equal tag tuples. Each action
segment gets one tag set per arm, and actions with equal tag sets form a
cluster [arxiv:2412.20998 §5.2] [§5.3].

```pseudocode
# Group action segments whose full tag sets are equal  [arxiv:2412.20998 §5.3]
# Input:  actions: list of (action_id, code), code = tuple of tags per arm
# Output: clusters: list of lists of action IDs, each with more than one member

def cluster_actions(actions):
    # Equal tag sets form the groups; keep only groups with more than one member
    groups = {}
    for action_id, code in actions:
        groups.setdefault(code, []).append(action_id)
    return [ids for ids in groups.values() if len(ids) > 1]
```

Bimanual tasks use one tag set per arm, and a unimanual task sets the second
arm to a none tag [arxiv:2412.20998 §4.6] [§5.2]. The tag letters in the
paper's table differ from its text in places, so map them before use
[arxiv:2412.20998 §4.2] [tab. 2].

## Contradictions

None known between these two cards. They count different things, a tag set
and a DoF count, so their numbers are not comparable.

## Papers

- [DexGrip](/papers/2411.17124): a soft gripper with an active palm and belt surfaces, with a DoF comparison across six grippers.
- [T-DOM](/papers/2412.20998): a tag set for deformable-object manipulation, applied by hand to ten tasks and compared with two older taxonomies.