The ADAPT competition

The ADvanced Agile ProducTion (ADAPT) competition aims at addressing typical dexterous manipulation tasks (e.g. heap sorting, picking of parts and precision placement) of industrial components involving intuitive, multi-modal interfaces and human communication channels (speech, gestures, gaze, etc.), involved in the assembly process of an industrial mechanical system.

The main functionalities tested in separate benchmarks (FBM) include the detection and classification of parts, estimation of part poses, and quality control of an assembly. The task-based benchmarks (TBM) address collaborative programming for assembly as well as the collaborative assembly of complex parts. Each benchmark is oriented towards industry-relevant problems. For example, parts feature reflective surfaces and unstable resting poses (e.g. cylindrical shapes).

The ADAPT competition is divided into two campaigns:

Cascade campaign is an online competition with data and targets only the functionality benchmarks, such as object detection, pose estimation and quality control by utilizing visual sensors.

Field campaignis a physical competition with real robots and targets task benchmarks such as collaborative robot programming and coordination.

ADAPT Campaigns

1st Field Campaign

During the campaign participants were able to evaluate the performance of your robot system and algorithms, in the field of agile production, with respect to assembly and human-robot collaboration. ADAPT benchmarks evaluate robot perception and actions (e.g., object detection, pose estimation, grasping and manipulation and collaborative robot programming). Objects used are 3D printed gear assemblies to be manipulated by any robot and any gripper.

Dry-run Field Campaign

Preparations for the ADAPT competition are active in the form of dry-runs. More info available at the campaign's webpage.

Dry-run Cascade Campaign

CEA LIST is coordinating the Sim2Real Challenge under ADAPT Competition. In this challenge, the task was to detect and recognize different parts of a gear assembly system and to estimate their position and orientation. The goal here is to assess the performance of detection, classification and pose-estimation algorithms that have limited access to real data during the training phase.


The ADAPT evaluation plan available below describes the competition and its evaluation and will be continuously updated. Additional information will be posted here as well.


Coordination team:

Roel Pieters (Tampere University, Finland)
Max Pfingsthorn (OFFIS, Germany)
Pierre Loonis (Proxinnov, France)
Farzam Ranjbaran  (CEA, France)