New Opportunities and New Challenges in Scaling Machine Learning in Industrial Robotics

Original

For decades robots have been essential to highly repetitive manufacturing industries such as the automotive industry, freeing human workers from repetitive and dangerous tasks. However, in recent years the stage has been set for a dramatic shift in the entire robotics paradigm. While robotic systems were previously benchmarked on sub-millimetre motor reliability over millions of cycles, today’s cutting-edge robotic systems are measured by their flexibility, robustness, ease of deployment, and scalability. To understand the rapid developments that are occurring at the intersection of machine learning and industrial robotics, it’s helpful to examine the need for such developments, the innovations that academics and companies are pushing to meet these needs, and the challenges that still remain.

The Need

The automotive and electronics industries were the top two integrators of robotic systems until 2016. At that point they encountered a rival: new automation needs driven by the rise of e-commerce fulfilment centres. In a consumer-driven market with a wide range of products (or SKUs) and shorter product life cycles, robotic systems were needed that could adapt quickly and scale to millions of products. E-commerce fulfilment and warehouse management sectors are now generating the majority of growth in demand for automation, and this trend is expected to continue.

New use cases aren’t the only driver of new needs. At exactly the time new demand for labour is being created, the traditional supply - human workers - is becoming more valuable and scarce. As ageing societies create labour shortages, industries from e-commerce to food manufacturing are exploring collaborative robotic systems that work symbiotically with and augment their human counterparts. As human/robot coworking environments become more common, roboticists will have to approach safety and reliability of autonomous system robots differently compared to traditional isolated robotic cells. Technologies such as machine learning and deep learning are providing robots with the autonomy required to recognize their environments and define their tasks.

The Innovations

In the race to satisfy these market demands and increase society’s productivity, both academia and private enterprise are developing new techniques to teach robots scene recognition and manipulation. On the academic front, one high-profile example of this type of research is the work of the UC Berkeley Robot Learning Lab. In 2017, Pieter Abbeel and his Berkeley team used virtual reality to teach a robot named “Brett” to pick up an apple and place it on a paper plate. The process involved a human teacher using VR gear to, in a sense, “hijack” the body of the robot and manually control its limbs to carry out the task. After carrying out the human motion several times with this method, the robot was able to achieve the goal autonomously even when the system variables were changed.

Meanwhile, at some of the world’s largest companies, researchers have also attempted to “teach” robots through a variety of demonstration modes. Researchers at Google have used relatively unstructured video of humans to create a form of unsupervised learning; they presented videos recorded from multiple viewpoints to their robots, enabling them to imitate object interactions and human poses. To further add to the mix, private non-profit labs such as OpenAI are training deep neural networks on object localization using simulated RGB images, which in turn allows their robots to identify objects in real-world scenarios. For-profit companies like Osaro (where the author currently works) are also doing research in this area. Osaro’s integrated perception and control platform utilizes simulation for motion planning (an approach that involves interfacing with several simulation engines including OpenRave and Unity) as well as synthetic data, which is used to train models thousands of miles from our offices.

The Challenges

Despite these innovations, autonomous systems present plenty of challenges for today’s roboticists. Take speed, for example. Because robots operate in dynamic environments in which physical objects obstruct their path, designers must develop planning algorithms that can replan the motion while still meeting time constraints. Human workers in fulfilment warehouses pick items at a speed of about 5-10 seconds per item. Matching this capability on machines is still a huge challenge that requires fine-tuning computations within algorithms and optimizing movement. Reliability is also an important area for improvement. Even today’s most sophisticated models are prone to dropping and mis-classifying objects. Achieving human level reliability will depend on better motion planning and control strategies. Lastly, the industry is grappling with the issue of scalability. Systems (such as end-to-end deep reinforcement learning algorithms) that function well in games and simulation are not necessarily equipped to handle industrial-scale workloads. On top of creating technology that overcomes all of the obstacles above, designers have to be sure that their products are adaptable for future growth and clients’ evolving needs.

On top of all these challenges, any solution to the challenges above must be economical.  A common method to address these issues is to fit systems with expensive sensors and other custom hardware. However, Osaro’s unique approach keeps hardware costs low by leveraging intelligent software. Our motion planning engineers integrate novel real-time inverse kinematics algorithms, joint-level path planning, and motion controllers to achieve smooth and collision-free paths that mimic human-level manoeuvrability. Going forward, the exponentially growing e-commerce industry will continue to challenge and motivate designers to find innovative techniques that address the issues of speed, reliability and scalability.  

About the Author

Hariharan Ananthanarayanan is a motion planning engineer at Osaro with over seven years of experience in the automation/robotics industry. Hari focuses on motion control and programming robotic manipulators using adaptive and torque based MIMO control, as well as collision-free path planning algorithms customized for a variety of environments and adapted specifically for combination with fast inference machine learning algorithms. Osaro is a San Francisco based machine learning company offering integrated perception and control software for industrial scale robotic deployments (ASRS systems, auto manufacturing, food prep, e-commerce, etc). Osaro began deploying its first solution OsaroPick, which integrates with e-commerce automated storage and retrieval systems, in Japan in early 2018.

Original

Industrial Automation Deep Learning Robotics


0 Comments

    As Featured In

    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original
    Original

    Partners & Attendees

    Intel.001
    Nvidia.001
    Acc1.001
    Ibm watson health 3.001
    Rbc research.001
    Mit tech review.001
    Facebook.001
    Graphcoreai.001
    Maluuba 2017.001
    Twentybn.001
    Forbes.001
    Kd nuggets.001
    This website uses cookies to ensure you get the best experience. Learn more