Machines learn common sense

Researchers from DARPA’s Machine Common Sense (MCS) program recently demonstrated a series of improvements in robotic system performance during several experiments. Just as infants must learn from experience, MCS seeks to build computational models that mimic core domains of infant cognition for objects (intuitive physics), agents (intentional actors), and places (spatial navigation ).

Using only simulated training, recent MCS experiments have demonstrated advances in systems capabilities – from understanding how to grasp objects and adapting to obstacles, to modifying speed/gait for various Goals.

“These experiments are important steps that bring us closer to building and fielding robust robotic systems with generalized motion capabilities,” said Dr. Howard Shrobe, MCS program manager in the Office of Innovation at the information from DARPA. “Prototype systems don’t need large suites of sensors to deal with unexpected situations that might arise in the real world.”

Adapt quickly to changing terrain

In one experiment, researchers from the University of California, Berkeley have developed a rapid motor adaptation (RMA) algorithm that allows quadruped robots to quickly adapt to changes in terrain. Using the RMA algorithm and proprioceptive feedback (the body’s sense of self-movement and position), the robots successfully navigated through a range of real and simulated terrain.

The algorithm is fully simulation trained without using any domain awareness type reference trajectories or predefined foot trajectory generators and is deployed without any fine tuning. Real-time terrain adaptation is essential for quadruped robots to help military units carry and sense the load.

Carry dynamic loads

Researchers at Oregon State have demonstrated the ability of a bipedal robot to learn to carry dynamic loads with only proprioceptive feedback. The robot, known as Cassie, learned common sense behaviors in a simulated-to-real learning environment. Cassie has adapted her approach to account for changes in load dynamics, such as liquid sloshing or balancing weights.

After simulation training, Cassie was able to walk on a treadmill for several minutes with four different types of dynamic loads. In contrast, before the scholarly common sense training, Cassie fell off immediately.

Understand how to grab objects

In natural environments, humans encounter a wide variety of possible tools, variations of tools, and objects. This variety presents a challenge for robots. They must foresee all the possibilities of functioning, which is why it is important that they have a general prehension capacity rather than a specialized capacity, for a predefined set of objects.

Researchers from the University of Utah, members of the MCS team at Oregon State University, have developed an active learning algorithm that allows robots with multi-fingered hands to dexterously grasp previously invisible objects when fully trained in simulation.

The new approach allowed the robot to grasp with more than 93% real-world success on new objects, compared to 78% of existing passive learning approaches.

Additional research

Another technical area within MCS seeks to develop computational tools that learn from reading the web, such as a research librarian, to build a repository of common-sense knowledge capable of answering natural language and evidence-based questions. images on common-sense phenomena.

MCS researchers from the University of Washington and two teams from the University of Southern California, Information Sciences Institute are currently using a variety of approaches, including hyperbolic learning. This technique learns the common sense structure of human behavior and physics from large collections of videos to predict human actions up to 30 seconds into the future.

Researchers are also building a scalable, machine-created symbolic knowledge base that will provide a better, broader, and more diverse representation of the world.

“By focusing on common sense, we create the ability for systems to have the flexibility of human learning and the breadth of human knowledge,” Shrobe said. “Merging this knowledge with advanced robotics could result in high-performance, mission-critical systems that humans will want to have as partners.”

Comments are closed.