Where the Intermediate is the Big Step Intralogistics with Safe and - PowerPoint PPT Presentation

Where the Intermediate is the Big Step Intralogistics with Safe and Scalable Fleets of Autonomously Operating Vehicles in Shared Spaces Achim J. Lilienthal contact: www.mrolab.eu/achim-lilienthal achim.lilienthal@oru.se

Intra- Logistics with Integrated Automatic Deployment Where the Intermediate is the Big Step Intralogistics with Safe and Scalable Fleets of Autonomously Operating Vehicles in Shared Spaces H2020 project funded by the EC Start: Jan 2017 www.iliad-project.eu Duration: 48 months Funding: 7M€ Partners: 9

Intra- Logistics with Integrated Automatic Deployment Where the Intermediate is the Big Step Intralogistics with Safe and Scalable Fleets of Autonomously Operating Vehicles in Shared Spaces H2020 project funded by the EC Start: Jan 2017 www.iliad-project.eu Duration: 48 months Funding: 7M€ Partners: 9

ILIAD in a nutshell Why the intermediate is the big step

Concept image More quickly changing market needs, unforeseeable trends, shorter product life cycles, …

Flexible intralogistics needed • Today’s highly automated goods-to-man solutions • require dedicated warehouses, • are unsuitable for fresh food, bulky goods, etc. • ILIAD’s aim is to provide solutions for flexible intralogistics for the transition to automation in shared spaces.

Flexible intralogistics needed Therefore we need intralogistic systems that are F-RE-SE–QUD–SA-EFF! (1) highly flexible, (2) rock-solid reliable, (3) self-optimising, (4) quickly deployable and (5) safe yet (6) efficient in environments shared with humans.

Approach Safety and efficiency by Easy deployment with long-term learning and semantic mapping prediction of patterns Flexible manipulation Human safety-aware AGV fleets: On-line, self-optimising safe and legible motion planning, fleet management human detection and tracking, mutual communication of intent

Platform (CitiTruck) Early prototype (APPLE platform at Hannover fair 2015)

Demonstrators

Demonstrators • Demos at key stakeholders from food distribution sector. • Distribution of fresh food products particularly challenging: • sensitive products, • short shelf life, • rapid response to consumer needs. • Food industry largest manufacturing sector in EU: 4.2 million jobs.

Orkla Foods Not always well-defined aisle environment Heterogeneous and brittle (and sometimes large or heavy) items to pick.

NCFM • National Centre for Food Manufacturing

ASDA Photo credit: Adrian Welsh Photo credit: Redirack

Easy deployment, long-term operation Learning, modelling and exploiting flow information

Learn, model and use dynamics • Learn and map how things usually move . • Statistical multi-modal flow model. • Use for socially compliant motion planning: less "annoying", safer and more efficient operation.

Learn and model dynamics • Learn and map how things usually move . • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. FP7 EU project SPENCER http://www.spencer.eu/

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Using this information can • lead to safer and socially more acceptable robot trajectories; • allow to plan energy efficient paths for flying robots; • improve gas distribution mapping.

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map (Circular–Linear Flow Field map) – A probabilistic approach for general flow mapping Enabling Flow Awareness for Mobile Robots in Partially Observable Environments. T. P. Kucner, M. Magnusson, E. Schaffernicht, V. Hernandez Bennetts, and A. J. Lilienthal. RA-L 2017 ( 2 :2, pp. 1093-1100) / ICRA 2017

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Local elements are probability distributions of observations V = ( θ , ρ ) • One circular (orientation θ ) and one linear (speed ρ ) random variable.

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. Learning CLiFF maps • Statistical multi-modal flow model CLiFF map • Init: Mean Shift (MS) to determine number of clusters and their positions • Local elements are probability distributions of observations ( θ , ρ ) • Use MS clusters to initialize Expectation Maximisation (EM) • One circular (orientation θ ) and one linear (speed ρ ) random variable. • Semi-wrapped Gaussian mixtures

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Local elements are probability distributions of observations ( θ , ρ ) • One circular (orientation θ ) and one linear (speed ρ ) random variable. • Field of semi-wrapped Gaussian mixtures

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Local elements are probability distributions of observations ( θ , ρ ) • One circular (orientation θ ) and one linear (speed ρ ) random variable. • Field of semi-wrapped Gaussian mixtures

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Local elements are probability distributions of observations ( θ , ρ ) • One circular (orientation θ ) and one linear (speed ρ ) random variable. • Field of semi-wrapped Gaussian mixtures

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Field of semi-wrapped Gaussian mixtures • Observations in robotics are often sparse ?

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Field of semi-wrapped Gaussian mixtures • Observations in robotics are often sparse ?

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Field of semi-wrapped Gaussian mixtures • Observations in robotics are often sparse • => Data imputation to build dense maps from sparse measurements • Monte Carlo Imputation (MC) • sampling virtual observations from the surrounding • tends to preserve multimodal characteristics and keep sharp transitions • Nadaraya Watson Imputation (NW) • Weighted extrapolation (distance kernel) • smooths data and models introduces gradual changes • MC performed better than NW on pedestrian data (and is less sensitive to kernel size) – but results may differ in different applications

Learn and model dynamics • Learn and map how things usually move. • Extensive research on mapping geometric structure. • Environment typically defined by spatial movement patterns. • Statistical multi-modal flow model CLiFF map • Field of semi-wrapped Gaussian mixtures • Observations in robotics are often sparse • => Data imputation to build dense maps from sparse measurements • Summary • It is possible to accurately represent multimodal (even turbulent) flow using CLiFF map. • It is possible to reconstruct a dense representation based on sparsely distributed observations.

Using dynamics models • Use for socially compliant motion planning: less "annoying", safer and more efficient operation. Kinodynamic Motion Planning on Gaussian Mixture Fields. L. Palmieri, T. Kucner, M. Magnusson, A. J. Lilienthal, and K. O. Arras ICRA 2017

Using dynamics models • Socially compliant motion planning using CLiFF map • Mobile robot motion planning approach: CLiFF-RRT* • CLiFF map – provides learned perception prior • RRT* motion planner • asymptotically optimal sampling-based motion planner • considers the robot's kinematic and its non-holonomic constraints Kinodynamic Motion Planning on Gaussian Mixture Fields. L. Palmieri, T. Kucner, M. Magnusson, A. J. Lilienthal, and K. O. Arras ICRA 2017