OpenAI Robotics Symposium 2019

OpenAI hosted its first Robotics Symposium on April 27, 2019, to create a multidisciplinary platform for exchanging ideas on building complex robot systems. The event focused on the goal of creating robots that can learn, which OpenAI identifies as the primary path toward integrating robots into everyday life.

Technical Approaches to Robot Learning

Several researchers presented diverse methodologies for enabling robots to acquire skills and adapt to their environments:

Sim2Real and Reinforcement Learning

Wojciech Zaremba (OpenAI) presented research on "Learning Dexterity," which utilizes sim2real with domain randomization and large-scale reinforcement learning with memory-augmented policies. This approach enables meta-learning, allowing a policy to transfer to a physical robot without requiring training on the physical hardware itself.

Self-Supervision through Play

Pierre Sermanet (Google Brain) described a method where "play" provides self-supervision for representation learning. This allows robots to acquire a diverse set of skills that can be recombined to solve novel tasks without the need for explicit labels or rewards.

Engineering vs. Deployment Learning

Leslie Kaelbling (MIT) discussed an architecture for building intelligent robots that distinguishes between learning in the "robot factory" (engineering time) and learning "in the wild" (deployment time) to facilitate the acquisition of new skills.

Runtime Adaptation

Jeff Clune (Uber AI / University of Wyoming) detailed work published in Nature regarding robots that can rapidly adapt at runtime if they sustain damage, increasing their resilience to unexpected environmental conditions or physical failure.

Human-Robot Interaction (HRI)

Research presented at the symposium focused on the 경로 (path) to making robots safe and socially intuitive:

Predicting Human Behavior

Anca Dragan (UC Berkeley) proposed a framework for predicting human actions that moves beyond the assumption that humans are "noisy-rational." Her framework accounts for strategic human behavior and systematic sub-optimality, such as a human's lack of knowledge regarding environment physics or evolving preferences.

Social-Emotional Intelligence

Jin Joo Lee (MIT / Amazon) explored the use of social-emotional intelligence to make robots more lifelike. This involves robots that can interpret human emotional expressions and communicate using similar emotional cues.

Industry Gaps and Practical Application

Chris Atkeson (CMU) provided a critical analysis of the gap between robot learning research and actual robot programming practice, discussing the requirements necessary to make learning robots truly useful in practical applications.

Dexterity Demonstration

As part of the event, OpenAI performed a live demonstration of a humanoid robot hand manipulating a block. The system utilized vision and reinforcement learning to achieve the task.

Symposium Outcomes

OpenAI reported that the symposium resulted in new internal perspectives on the tension between traditional robotics and deep learning robotics, as well as new ideas regarding self-supervision. The organization intends to repeat the format, potentially as an annual event, to foster a shared understanding of the diversity of approaches to solving robotics problems.