- Marc Raibert’s new institute aims to make robots like Spot and Atlas more capable of independent thinking and action.
- The institute will leverage recent AI advances like large language models and reinforcement learning to push robot intelligence forward.
- Early projects include repairing bicycles, controlling an agile two-wheeled vehicle, and improving robotic manipulation.
- Raibert believes the public responds positively to engaging with advanced robots like Spot despite some fearful reactions.
- He is surprised by warnings from AI leaders about risks from intelligent robots, arguing generalized robotic intelligence is still a long way off.
- Raibert remains dedicated to advancing robot capabilities into new challenging environments through his institute’s research.
Marc Raibert, founder and executive director of the Boston Dynamics AI Institute, has built some of the most advanced robots in the world during his time at Boston Dynamics. Acquired by Hyundai in 2020, Boston Dynamics developed agile legged robots like Big Dog, Spot, and the humanoid Atlas. While Raibert’s robots can dance and do parkour thanks to intricate programming by humans, his new institute aims to make them more autonomous.
Teaching Robots to Be More Independent
The Boston Dynamics AI Institute, launched in August 2023 with Hyundai’s backing, will research ways for robots to comprehend and tackle unpredictable situations with little or no human assistance. Raibert believes recent advances in AI, like large language models, could aid this goal. One area the institute will explore is how language models that have absorbed vast amounts of textual data from humans can be applied to robotics. While not focused on physical acts, the models contain some information about human embodiment that could prove useful.
Raibert also plans to apply reinforcement learning techniques used extensively in simulation to physical robots. As an example, he cites work by Marco Hutter at ETH Zurich, who will collaborate with the institute. Hutter has used reinforcement learning for robots to teach themselves climbing skills in simulation. His robots can then transfer those learnings to climb real objects they have never encountered before using various parts of their bodies.
The institute will focus on several challenging robotics projects requiring intelligence and perception. One group is working on programming robots to repair bicycles autonomously. While less complex than fixing cars, bicycles still pose difficulties compared to straightforward warehouse tasks. The team currently has a robot in the lab attempting bicycle repairs, but it remains in early stages.
Another ambitious project involves building a two-wheeled vehicle with jumping and bouncing capabilities. Inspired by daring parkour cyclists, the robot bike will need strong vision systems to spot objects and plan sequences of jumps. This mimics how human cyclists understand their surroundings and plot out creative routes.
In addition, the institute will advance research into dexterous robotic manipulation—an area Raibert believes has lagged behind mobility advances. Robots still struggle with many manipulation tasks that humans perform effortlessly. By combining improved hardware, software, and learning approaches, the researchers hope to close this gap.
While public reactions to advanced robots are sometimes fearful, Raibert believes people engage positively with his robots. He cites how crowds love to interact with and take selfies with the Spot robot. Even individuals who claim to fear robots seem to make exceptions for certain designs. Raibert questions whether the worries about robots are overblown compared to more pressing human-caused harms.
Nonetheless, Raibert is surprised by warnings from AI leaders about risks from intelligent robots. He argues generalized robotic intelligence that could threaten human roles is still a distant prospect. In most cases, robots only aim to match narrow human skills for specific jobs. While robot capabilities may someday advance drastically, Raibert believes that breakthrough is a long way off even given recent AI progress.
Raibert remains dedicated to pushing robot capabilities forward into new environments through his institute’s research, arguing mobility is only one piece of the puzzle. By tackling complex real-world robotics problems, he hopes to close the gap between robots’ physical and cognitive abilities. Even if the road ahead is long, Raibert is committed to the quest for robotic autonomy.
Two products launched right now:
AI Agent: https://orbitmoonalpha.com/shop/ai-tool-agent/
AI Drawsth: https://orbitmoonalpha.com/shop/ai-tool-drawsth/