Marc Raibert on Aggressive Robotics, Spot, and AI Institute

Marc Raibert, a legendary roboticist and founder of Boston Dynamics, recently discussed his journey in robotics, the philosophy behind Boston Dynamics' creations, and the vision for the newly established Boston Dynamics AI Institute.

Early Fascination with Robotics

Raibert's interest in building began at a young age, influenced by his father's passion for engineering. However, his true calling to robotics solidified in 1974 during graduate school at MIT. While studying neurophysiology in the brain and cognitive sciences department, he found himself drawn to the AI lab. A pivotal moment occurred when he saw a robot arm disassembled into a thousand pieces. This sight ignited his imagination, leading him to realize his path as a roboticist. He felt that neurophysiology wasn't conceptual enough to understand control systems or thought, and the AI lab offered a more appealing avenue.

He noted the historical tension and eventual bridging between brain and cognitive sciences and robotics, citing figures like David Maher who created models appealing to both biologists and computer scientists.

The Philosophy of Aggressive Robotics

Raibert's work is characterized by robots that are not cautious, a stark contrast to many early robots designed to move slowly and carefully. This philosophy stemmed from an observation at a biological legged locomotion conference where he saw a six-legged robot moving very slowly, always maintaining tripod stability. He immediately recognized this as fundamentally different from how humans and animals walk, which involves bouncing, flying, and utilizing the springiness of legs.

Inspired to try the opposite approach, Raibert embarked on creating hopping robots. Initially, he focused on the energy of bouncing, with the concept of balance in 3D coming later. This led to the development of one-legged, pogo stick-like robots. He believes a similar "aggressive" approach is needed in robot manipulation, moving beyond static grasping to more dynamic, human-like interactions such as juggling or sorting multiple objects. While this approach might lead to initial failures, he sees it as a long-term strategy for achieving human-level dexterity. He referenced Matt Mason's analysis of Julia Child, who used numerous non-grasping techniques to manipulate objects, highlighting the complexity and dynamism of human interaction with the world.

The Legendary Leg Lab and the First Hopping Robot

The Leg Lab, which Raibert founded, began at Carnegie Mellon in 1986. The first hopping machines were developed there, with a simplified version working around 1982 and a 3D version in 1983. The Quadruped, a more advanced robot, was built in 1984 or 1985 and became fully operational around 1986 after years of development.

Raibert recounted the origin of the first hopping robot, which began during his time at JPL. Encouraged by Ivan Sutherland, a pioneer in computer graphics, Raibert proposed a "pogo stick robot" project. Sutherland provided $3,000 for the initial model, which Raibert built himself. They then presented this prototype to Craig Fields at DARPA, securing $250,000 in funding, a significant amount for research in 1980. At the time, Raibert didn't envision the practical applications of such robots, focusing instead on academic progress and understanding the fundamentals of animal locomotion.

Technical Challenges of the Hopping Robot

Creating the hopping robot involved significant engineering. Raibert collaborated with Ben Brown to refine the design and actuation system. The initial setup involved floating the robot on an inclined air table, then transitioning to a device that allowed it to run around a room, and finally, making it work in 3D.

The balance mechanism of a hopping stick in 3D involves three main components: 1. Bouncing: A system estimates the robot's height to control energy input, ensuring consistent hops. 2. Foot Placement: Calculations determine where to place the foot relative to the center of mass to maintain balance, similar to a pole vaulter. 3. Attitude Control: Torque is applied between the legs and body while the foot is on the ground to keep the robot upright and prevent uncontrolled rotation.

Early robots were not optimized for extreme performance, but subsequent work by his students focused on increasing speed and navigating obstacles.

The Birth of Boston Dynamics and BigDog

Raibert founded Boston Dynamics in 1992. Initially, the company focused on physics-based simulation. A turning point came when they quietly worked on a running version of Sony's Aibo robot, which helped them rediscover their passion for robotics. This experience, along with developing tools for Sony's QRIO humanoid robot, brought them back to their core mission.

One early project at Boston Dynamics was a surgical simulator with force feedback, designed to train surgeons. While technically fascinating and well-received by surgeons, it proved commercially unviable due to differing expectations about payment. This led to a pivotal decision to abandon the simulator and fully commit to building robots.

The first significant robot under the Boston Dynamics flag was BigDog, a large quadruped developed as part of a DARPA biodynamics program. BigDog was revolutionary because it integrated all power and computing onboard, unlike previous Leg Lab robots that required external hydraulic pumps and computers. It used a gasoline engine to power hydraulic actuation, presenting complex engineering challenges.

BigDog was extensively tested on the Guadalcanal Trail at the Marine Corps base in Quantico. These real-world tests highlighted the challenges of balancing a robot on uneven terrain. While early versions required a human operator for visual perception, later electric prototypes (precursors to Spot) demonstrated significant advancements in controls, allowing amateurs to operate them effectively.

BigDog evolved into LS3, a load-carrying robot designed to carry 400 pounds but capable of carrying up to 1,000 pounds. LS3 could travel 20 miles on gasoline power.

From Hydraulics to Electric: The Genesis of Spot

The transition from BigDog/LS3's hydraulic power to electric power was a significant step. Larry Page, then at Google, challenged Raibert to create a smaller, less intimidating robot, leading to the development of Spot. This involved moving from hydraulic actuation to an all-electric, non-hydraulic design.

Raibert expressed his continued appreciation for hydraulics, noting its performance advantages in terms of strength-to-weight ratio. Boston Dynamics innovated hydraulic systems, designing smaller, more efficient valves and integrated power supplies. However, the move to electric power for Spot was driven by the need for a less intimidating, more house-friendly robot.

The Art of Natural Movement

The natural and beautiful movement of Boston Dynamics robots is a hallmark of their design. Raibert attributes this to: * Good Hardware: He emphasizes that hardware innovation is still crucial. * Dynamic Approach: Instead of purely reactive servoing, the robots use a predictive approach, anticipating future motion and adjusting accordingly. For Spot, this involves a limited horizon calculation of a couple of seconds, constantly iterating. * Somersaults and Complex Maneuvers: For complex actions like somersaults, the robot must plan much further ahead, coordinating momentum and rotation for a successful landing. The first planar somersaulting robot was built in the mid-1980s. Later, a two-legged robot performed a 3D somersault, requiring additional techniques like tucking legs to increase rotation rate. Raibert noted that humans often perform such complex movements without fully understanding the underlying physics, highlighting how robotics can illuminate human biomechanics.

Raibert also mentioned Wildcat, a quadruped that achieved 19 miles per hour on flat terrain, showcasing the company's pursuit of speed.

The Role of Knees and Passive Dynamics

The inclusion of knees in legged robots, starting with BigDog, marked a shift. Human knees involve complex musculature and the ability to store negative work. While BigDog had pogo stick springs for compliance, later robots like Spot moved towards more energy-driven controls.

Raibert discussed the concept of "passive dynamics," where a mechanical system can generate motion without constant computer control, as demonstrated by Tad McGeer's work on legged systems that could walk down an incline. This highlights that the body itself is an active participant in motion, not just a recipient of commands from the "brain." A well-designed robot incorporates this mechanical efficiency.

Boston Dynamics AI Institute: Combining Athletic and Cognitive Intelligence

Raibert is now leading the Boston Dynamics AI Institute, which aims to combine the "athletic intelligence" (physicality, mechanical design, real-time control) that Boston Dynamics is known for with "cognitive intelligence" (planning, understanding, learning).

The institute's vision is to create smarter robots that can: * Learn from Observation: Watch a human perform a task, understand it, and then execute the task themselves (on-the-job training for robots). This is a long-term, "science fiction" goal. * Inspect, Diagnose, and Fix: Robots that can examine machines, identify problems, and perform repairs, building on existing capabilities like data collection for machine health.

The institute employs a "stepping stones to moonshots" approach, breaking down ambitious goals into smaller, tangible milestones to provide feedback and motivation.

Machine Learning and the Future of Robotics

Machine learning plays a significant role in the AI Institute's work, especially with the rapid advancements in large language models. However, Raibert notes that applying machine learning to physical robots differs from language processing, as "pixel values aren't like words." The institute has a strong machine learning component, including an office in Zurich led by Marco Hutter, a leader in reinforcement learning for robots.

While traditional control methods like model predictive control still drive impressive athletic performances (e.g., Atlas's movements), Raibert believes the future lies in a "mating of the two" approaches.

Building Great Teams: Technical Fearlessness, Diligence, Intrepidness, and Fun

Raibert outlined four key components for building a great team: 1. Technical Fearlessness: Willingness to tackle problems without knowing the solution, finding entry points, and learning from iterative progress. 2. Diligence: A commitment to developing robust solutions that can handle variations in tasks and environments, as demonstrated by Boston Dynamics' rigorous testing, including intentionally perturbing robots during tasks. 3. Intrepidness: Perseverance in the face of failure, understanding that robotics is challenging and requires continuous effort. He cited Atlas's 109 attempts to climb three steps as an example, emphasizing the importance of robust hardware and a willingness to repair. 4. Technical Fun: The satisfaction derived from engineering, which combines scientific rigor with creative artistry. He highlighted the joy of building, the potential for impact, the collaborative nature of teamwork, and the financial rewards.

Competition and the Future of Robotics

Raibert expressed admiration for Elon Musk's technological achievements with Tesla and SpaceX, and acknowledged the potential of Tesla's Optimus robot, even if it's not yet at Atlas's level. He hopes for a "robot meetup" between Atlas and Optimus.

He noted that the AI Institute works with a variety of robots, including Spots, ANYmal robots, and various robotic arms, creating a "robot playground" for research.

Regarding competition, Raibert stated that for many years at Boston Dynamics, they didn't focus on it, as they were creating a new category. However, in the cognitive AI space, competition for talent and resources is more intense. He believes that in the quadruped market, the presence of multiple companies is beneficial, shifting the user's question from "Can a quadruped do my job?" to "Which quadruped do I want?" He anticipates a similar evolution for humanoid robots.

He is optimistic about the potential for significant cost reduction in robots through mass production and engineering innovation, citing Hyundai's expertise in manufacturing as a key factor for Boston Dynamics.

Social Robotics and the Meaning of Life

Raibert believes that social robotics, with robots in homes, will eventually become a significant market, driven by the human desire for intelligent companions. He acknowledged the challenges of achieving performance, safety, and cost simultaneously, and the public's potential suspicion of robots with cameras in their homes. However, he drew parallels to smartphones, which are widely accepted despite privacy concerns.

On the topic of artificial general intelligence (AGI), Raibert expressed skepticism about the widespread fear surrounding it. He questioned why humans would feel threatened by a smarter computer, suggesting that such systems could still be under human control and beneficial, much like other technologies with inherent risks. He compared the AGI fear to the early concerns about nuclear chain reactions, viewing it as an opportunity rather than an existential threat.

Finally, Raibert shared his personal philosophy: "You have to have fun while you're here." He believes that engineering offers a unique blend of scientific exploration and artistic creation, allowing individuals to build things that didn't exist before and have a meaningful impact on the world. He finds the "metal to life" moment of robotics truly magical, especially the elegance and grace of movement that evokes life. He hopes to see robots dancing with humans, adapting to human movements, and even showcasing unique "robot animation" styles.

His advice to young people is to identify what they would do if there were no constraints and then strive to get as close to that ideal as possible, emphasizing that opportunities are often wider than perceived and that sustained pursuit leads to remarkable achievements.