Defining the Taxonomy for Industrial Dexterity

Industrial sites demand manipulation that goes beyond grasping. DexBench presents Action definitions that will set the standard for industrial robotics.

Actions
16
Core motions drawn from industrial processes
Skills
78
Units of ability that each resolve one bottleneck within an Action

Problem

The Gap Between Industry and Research

Dexterity research still lacks a common framework for defining and comparing Action difficulty across domains.

  1. 01

    No Shared Action Vocabulary

    Each industry describes its manipulation requirements in its own terms. A “pick” in logistics and a “pick” in electronics assembly share a word, but not a method.

  2. 02

    No Common Difficulty Scale

    The same Action name covers different conditions. Inserting a USB cable and inserting a piston rod are both “insertion,” yet they require different abilities.

  3. 03

    No Reproducible Conditions

    Different objects, different initial conditions, different success criteria. Results are hard to compare even between labs, and harder still between companies.

  4. 04

    No Clear Development Priorities

    Without a map that shows manipulation abilities at a glance, teams struggle to decide what to develop, invest in, or research.

Anatomy

Action · Skill · World

Comparing real-world tasks requires breaking them down into common units. DexBench uses three levels, from a single motion to a complete on-site task.

  • Action

    The smallest unit of manipulation that serves a single purpose

  • Skill

    A unit of ability that resolves an independent bottleneck within an Action

  • World

    A task from a real industrial site that can be broken down into Skills for analysis

Evaluation

DexBench - Simulation Benchmark

The DexBench evaluation platform provides Worlds modeled on real industrial sites across sectors, where every model is tested against the same criteria.

DexBench

Measuring dexterity where it matters.

  • Retail

    Drink from a fridge shelf

  • Logistics

    Box from cage to conveyor

  • Manufacturing

    Drawer into a washer

  • Hotel

    Plate stack to a cart

GENERALIZATION

  1. GA-O

    Object

    Variation in SKU, geometry, size, or appearance while preserving the task. Evaluation measures performance on unseen objects.

  2. GA-P

    Physics

    Variation in mass, friction, stiffness, center of mass, force/torque, and dynamic states such as speed and timing.

  3. GA-S

    Spatial

    Variation in initial pose, arrangement, clutter density, access constraints, clearance, and target position.

Framework

Classification Framework

Object State Complexity (OSC)

Six axes that classify why a Skill is difficult

C_geomC_forceC_contactC_obsC_deformC_dyn
View OSC Axes

Dexterity Regimes

Five capabilities needed to resolve the conditions identified by the OSC Axes

E1E2E3E4E5
View Dexterity Regimes

Actions

16 Actions of Industrial Value

Method

How We Define Skills

  1. 01

    Bottleneck-Based Selection

    Even within the same Action, the bottleneck shifts with the environment, the object, and the limits of what a model can do. From these, DexBench selected the bottlenecks with industrial value and defined each as a Skill.

  2. 02

    Generality

    A Skill is defined as an Action paired with a bottleneck. Tasks that share the same bottleneck count as the same Skill, even when the objects differ.

  3. 03

    State Transition

    DexBench defines what changes, not how it changes. A Skill specifies the initial and final states and leaves the method to the model.

  4. 04

    Framework Mapping

    Each bottleneck is difficult for a different reason and calls for different capabilities. Each Skill is therefore linked to the OSC axes that explain why it is difficult and to the Dexterity Regimes it requires.

Partners

Our Partners

Platform Partners

Industry Partners