Skip to content

Berkeley lab hopes to birth teachable robotic systems

By Suzy Marzano Senior Manager, Industry Development and Technical Activities, SME

A team of researchers from UC Berkeley, led by Pieter Abbeel, is working on the creation of smart robots that are teachable and can learn new skills without pre-programming. Abbeel and his team also formed a startup called Embodied Intelligence with the aim of developing artificial intelligence (AI) software to enable robots to learn from humans to perform complex tasks.

Another day at AI software firm Embodied Intelligence, another discussion of mathematical formulations. Left to right: research scientist Tianhao Zhang, Chief Scientist & President Pieter Abbeel, CTO Rocky Duan and CEO Peter Chen. The team also conducts robotics research at UC Berkeley.

There are many advantages to implementing robotic automation into the manufacturing of goods and products. Robots are well suited to performing repetitive tasks consis­tently over long periods of time, performing tasks that can be dangerous, and performing tasks that require precision.

Limitations to implementing robotic automation include tasks that require decision-making, flexibility, adaptation and job learning. Programming robots for complex tasks can also be costly and time-consuming.

Embodied Intelligence is working on bridging the big gap between the capabilities of humans and robotic manipulators.

Abbeel’s research at UC Berkeley’s Robot Learning Lab paved the way for Embodied Intelligence. Early accomplishments from that lab include teaching a robot to tie knots; reliably and autonomously fold previously unseen towels of different sizes, and screw a cap onto a water bottle.

In order for robots to perform complex tasks, researchers had to teach them how to adapt to changes in the environment. Most robots are programmed to handle specific tasks in a controlled environment and are programmed to handle each step of a task. Instead of programming robots to handle each step of a task, Abbeel’s team programmed them to observe and mimic humans demonstrating a task.

Embodied Intelligence will use machine-learning methods, including deep reinforcement learning, deep imitation learning, and “few-shot” learning. Deep learning is loosely inspired by neural networks of the human brain where we perceive and interact with the world. Imitation learning is a methodology whereby skills are acquired by demonstration observation. Few-shot learning is the ability of the robot to acquire new skills from a few demonstrations with the ultimate goal of new skill acquisition by one-shot or a single demonstration. Abbeel’s team has demonstrated how robot skill acquisition can be more efficient and effective with the use of virtual reality and tracking software.

The da Vinci Surgical System, from Intuitive Surgical, is an example of a teleoperation system designed for robotic manipulation whereby a surgeon’s hand movements are translated into smaller, precise movements of tiny instruments inside the body. However, the surgical system can be expensive and specialized hardware is needed.

Abbeel’s team has found that robotic manipulation through virtual reality teleoperation is more intuitive and natural at demonstrating and performing complex tasks.

VR Teleoperation enables the human operator and robot to perceive the same 3D environment. This allows for higher quality demonstrations and near elimination of human errors or visual distractors (i.e. humans making decisions based on information not available to the robot). The data collected is thus higher quality, which translates into the robots learning new skills faster.

“What we are seeing here instead is anyone who can use a VR headset can teach a robot new skills quickly,” said Peter Chen, co-founder and CEO of Embodied Intelligence.

A key to this plan is switching to the robotic system learning a skill rather than performing a sequence of events. Ideally, the idea could be extrapolated to adapt to changes and variables for a given task despite instructor demo variations and even mistakes.

More complex tasks will require more data and high-quality data. Researchers will need to determine how much data is needed to perform a given task.

The goal, Abbeel said, is to bring the cutting-edge AI research to robotics and manufacturing.

ACL repair innovation on tap at NC State

Researchers at NC State University are studying the manufacturing of new implantable medical products that can mimic the properties and function of soft tissues. The research surrounding tissue engineering and regenerative clinical approaches have shown promise to overcome current challenges of anterior cruciate ligament (ACL) and meniscus repair.

Currently, medical implants and products are being used in many different areas of the body and applications, such as: orthopaedics, cardiovascular pacemakers and defibrillators, neural prosthetics and medication delivery.

As people’s average life expectancy increases, the number of age-related diseases, such as degenerative joint diseases, increases. Osteoarthritis and rheumatoid arthritis, for example, can affect the structure of the movable (synovial) joints, with the hip, knee, shoulder, ankle and elbow. There are few effective ways to treat the root causes of the arthritic diseases, and in many cases clinicians manage symptoms only using medications or devices.

Implantable medical devices are central to the treatment of several arthritic conditions and can include total joint prostheses for arthritic hips and knees.

Most implantable medical devices are currently made of bioinert metal, polymers (plastics), and ceramics. Biodegradable and bioactive materials and living cells will soon replace bioinert metals and polymers to create living functional tissues and organs.

Rohan Shirwaiker’s “3D tissue manufacturing” research team is focusing on the design and scalable manufacturing technologies for engineered tissues. They are studying how the interactions between biomaterials and biofabrication processes affect the structural and functional characteristics of 3D scaffolds and bioprinted constructs. Based on an understanding of these material-process-structure-function relationships, they are developing new biofabrication approaches and tissue-engineered medical products, primarily in orthopaedics, in collaboration with other engineers and surgeons.

A key development for Shirwaiker included the ability to create an anatomically designed knee meniscus soft tissue construct made from a nanofiber-hydrogel bioink. Human adipose-derived stem cells were encapsulated in the bioink, and the material was bioprinted using extrusion-based additive manufacturing technology. His team’s research into the processing technique for the bioink and extrusion-based 3D bioprinting was detailed in a paper published in ACS Biomaterials Science and Engineering.

In another recent study, Shirwaiker’s team demonstrated how the porous micro-architecture of biodegradable 3D-printed polymer scaffolds can affect the cellular and extracellular matrix characteristics of engineered tissue using a rat model. This knowledge will help in creating biological substitutes to replace tissues, such as the knee meniscus and ACL.

Development is also underway on new biofabrication processes to create engineered tissues that more closely mimic the structural and functional properties of such orthopaedic soft tissues.

If the sensor fits …

Engineers from Florida State University’s High-Performance Materials Institute, in collaboration with scientists from Institut National des Sciences Appliquées in Lyon, France, have developed new sensor technology that could make wearable technology much more affordable to make and more sensitive to movement.

Wearable technology is on the rise—for use at home and in the workplace. A wearable device is essentially a smart platform with sensing, data processing, storage and communication capabilities. Many wearable devices also include interfaces and capabilities that provide feedback to the user.

Wearable technology is evolving at a rapid rate and is being applied to: clinical applications and monitoring health conditions, treatment for hearing-impaired people, physical fitness tracking and synchronization of data and communication. Many devices provide feedback to the user. For example, monitoring devices continuously measure and monitor blood glucose levels and transmit data to the user.

Wearable devices are rapidly advancing in terms of the capabilities, including functionality and size, with real-time applications.

Central to the collection of high-quality data is point of contact for the sensors. Today, most commercially viable sensors are created from metallic or semi-conductor materials. Semiconductor sensors are more sensitive to movement than metallic materials but are rigid and brittle. Metallic sensors are more flexible but are not as sensitive to movements and are more complex to make.

Engineers at The Florida A&M University—Florida State University College of Engineering, the joint engineering school of Florida State and Florida A&M University have developed a class of sensors that are more sensitive, flexible in structure, and less costly to make than currently available industrial sensors. The sensors are lower in profile, more sensitive to movements, more flexible and thus able to conform to surfaces on the body and cost-effective to make. The R&D, including properties of the motion sensors, was detailed in a paper published in Materials in Design.

The motion sensors were created using seven-micron-thin, flexible sheets of pure, durable carbon nanotubes called buckypaper. The novel sensors were printed on a commercially available ink-jet printer.

FSU’s High-Performance Materials Institute researchers detailed the sensors characteristics. After the motion sensors were integrated into a fabric glove, researchers performed sensitivity studies. They found that the novel sensors were more flexible and more sensitive to subtle movements and strains. Researchers are working to develop even thinner sensor material to allow for seamless integration into clothing or other wearable products.

Researchers take Kapton beyond thin sheet status

Researchers at Virginia Tech demonstrated a novel processing technique of a polymer, for what was thought to be a non-processible polymer. Researchers were able to 3D print a fully aromatic polyimide to create high-performance structures with significant strength and heat resistance.

The material, formally known as Kapton, is an aromatic polymer composed of carbons and hydrogens inside benzene rings, with exceptional thermal and chemical stability. Until now, engineers could only use Kapton as 2D-thin film.

Virginia Tech researchers created these complex 3D self-supporting Kapton parts, a lattice brick and VT logo. They were processed at 200 µm layers, 50 layers, and printed at 7.6 mm/s. The UV intensity was 20 mW/cm2 with wavelengths of 300–500 nm.

Kapton is often used in the multi-layer insulation that forms the outer wrapping of space instrumentation, including spacecraft and satellites, to protect the instruments from extreme heat and cold. Kapton tape is also used in the manufacturing of electronics to protect components during the production process.

Prior to the discovery, the polymeric structure could only be used in 2D applications. Virginia Tech researchers discovered a way to synthesize the polymers, opening up new ways for product designers and engineers to take advantage of Kapton’s exceptional material properties as 3D objects.

Timothy Long, a chemistry professor in the College of Science and Christopher Williams, a mechanical engineering professor, teamed up in the discovery.

Long also directs the Macromolecules Innovation Institute (MII). Williams is an associate director of MII and the director of the Design, Research, and Education for Additive Manufacturing Systems (DREAMS) Laboratory.

Long’s team of grad students developed the novel synthesis routine, and Williams’ team considered what could enable 3D printing.

Their joint research was detailed in a journal called Advanced Materials.

As the engineers began printing trials, the chemists tweaked the formulation to make the material compatible with printing systems.

The discovery of the new synthetic process for the polyimide paved the way for the material to be processed in AM to create high-performance structures in 3D.

One of the biggest limitations of materials currently being used in additive manufacturing processes is the loss of mechanical strength of the polymers above 300°F (149°C).

The newly synthesized polyimide exhibited stability across a very large temperature range and maintained its properties above 680° F (360°C). “We are now able to utilize the highest temperature polymer ever in an additive manufacturing process—about 285°F (141°C) higher in deflection temperature than any other existing printable polymer,” Williams said. “The material was also equivalent in strength to the conventionally processed thin-film Kapton material.” (The material’s heat-resistant ceiling before degradation is 1020° F [549°C]).

  • View All Articles
  • Connect With Us

Related Articles

Webinars, White Papers and More!

SME's Manufacturing Resource Center keeps you updated on all of the latest industry trends and information. Access unlimited FREE webinars, white papers, eBooks, case studies and reports now!