PFN in Warehouse Automation Then, now, and the future May 30, 2018 Wilson Ko
Who are we? Tokyo based company focused on Machine Learning and Distributed Computing, branch office in Berkeley We’re hiring at both locations! Major fields of interest • Autonomous Driving • Bio healthcare • Manufacturing www.toyota.ca www.bioprocessonline.com www.fanuc.co.jp 2 / 17
Brief history 2006 • Preferred Infrastructure founded 2014 • Preferred Networks founded 2014 • Partnered with Toyota Motor Corporation 2015 • Released Chainer 2015 • Partnered with Fanuc Corporation 2016 • Amazon Picking Challenge (2nd place) 2016 • Partnered with Japan National Cancer Center 2017 • Launched private supercomputer 2017 • Trained ImageNet in 15 minutes 3 / 17
Powered by Chainer • Flexible: Flexibility to represent new complex models • Intuitive: Suitable for rapid prototyping - Define by run • Powerful: Supports CUDA and multiple GPUs with little effort. • Open Source: https://github.com/chainer/chainer 4 / 17
Visual Inspection Detect anomalies on products, e.g. scratches, stains, bumps etc. Used to inspect car exteriors, smartphones, fabric etc. Challenges • Not much data available of damaged products • Detailed annotations of the anomaly location are needed Our solution • Automated defect detection system powered by Chainer • Only 100 normal pictures and 20 for defects to train 1 1 The number of pictures may change per target 5 / 17
Fanuc Bin-picking https://youtu.be/ydh_AdWZflA 6 / 17
APC 2016 7 / 17
Warehouse Bin-picking - APC 2016 8 / 17
Warehouse Bin-picking - APC 2016 • 200 msec inference time per scene 9 / 17
Warehouse Bin-picking - ICRA 2017 10 / 17
Warehouse Bin-picking - Latest system • Asynchronous planning/movement by decoupling camera • Average cycle time of 6 seconds vs 12 seconds for humans • Handling unknown objects - no registration needed 11 / 17
Warehouse Bin-picking - Latest system https://youtu.be/5pSq7qFAL0M https://youtu.be/x0IEHDy5G-8 12 / 17
Interactive Bin-picking Given spoken command, pick an item and place it at desired destination Challenges • A wide variety of object referring expressions ”a bear doll”, ”the animal plush”, ”that brown fluffy thing” • Human instructions are ambiguous and erroneous ”a dog doll?” 13 / 17
Interactive Bin-picking - Architecture Interactively Picking Real-World Objects with Unconstrained Spoken Language Instructions - ICRA 2018 14 / 17
Please come by! Interactively picking real-world objects with unconstrained spoken language instructions • Interactive poster session: Wed 10:30-13:00 @ pod T • Best award candidate on HRI: Wed 16:00-16:15 @ Great Hall https://youtu.be/DGJazkyw0Ws 15 / 17
Future work • Visual inspection • Keep improving UX • Ensure quality • Reduce data needed • Bin-picking • Improve picking accuracy • Multi-fingered grasping • Few-shot learning • Online learning 16 / 17
Please come by! We’re hiring in Tokyo and Berkeley! Interactively picking real-world objects with unconstrained spoken language instructions • Interactive poster session: Wed 10:30-13:00 @ pod T • Best award candidate on HRI: Wed 16:00-16:15 @ Great Hall 17 / 17
Recommend
More recommend
