Boston Dynamics, supported by South Korean carmaker Hyundai, is progressing in the development of its Atlas humanoid robot. The all-electric model, featuring an AI system powered by Nvidia microchips, has demonstrated advanced mobility and manipulation skills through a range of training methods, including supervised learning, motion capture, and simulation. The company is working towards deploying Atlas for tasks currently performed by human workers, while also acknowledging current limitations and long-term development timelines.
Atlas Development and Technical Specifications
Boston Dynamics, in which Hyundai holds an 88% stake, is advancing the development of its Atlas humanoid robot. The unit, measuring 5-foot-9-inches and weighing 200 pounds, has transitioned from an earlier hydraulic design in 2021 to an all-electric model. This current iteration integrates an AI system powered by Nvidia microchips.
A key engineering feature of the Atlas robot involves the omission of wires crossing the joints of its limbs, torso, and head. This design choice aims to facilitate continuous rotation and enhance reliability by reducing potential wire breakage.
Advanced Capabilities and Dexterity
Atlas has demonstrated a range of physical capabilities, including running, crawling, skipping, dancing, jumping jacks, and cartwheeling. The robot is also capable of self-righting from a prone position using only its feet. Its design allows for 360-degree rotation of its arms, head, and torso. Furthermore, Atlas can rotate its upper torso 180 degrees to change direction without reorienting its lower body.
For manipulation tasks, Atlas is equipped with three-digit hands designed for versatility. These digits can adjust to various configurations, functioning as a three-fingered hand or with one digit repositioned to act as a thumb. This design enables different grasp types, such as two-finger opposing grasps for small objects and wider grasps for larger items. Tactile sensors on the robot's fingers provide data to its neural network, aiding in object manipulation with appropriate pressure.
Demonstrations of Atlas's capabilities include practicing autonomous sorting of roof racks for an assembly line at Hyundai's Georgia factory, as well as tasks like stacking cups and tying knots.
Training Methodologies
The AI system of the Atlas robot is trained using several methodologies:
- Supervised Learning and Teleoperation: Machine learning scientists use virtual reality (VR) headsets to directly control the humanoid, guiding its movements until a task is completed. This process generates data for training the robot's AI models for autonomous task execution. Human operators also control the robot through teleoperation for specific tasks.
- Motion Capture: Human movements, such as jumping jacks, are recorded using motion capture body suits. This data is then fed into Boston Dynamics' machine learning process.
- Simulation: Over 4,000 digital Atlas avatars undergo training in simulation environments, often for durations such as six hours. These simulations introduce challenges like slippery floors or stiff joints to optimize the robot's performance of specific actions. The optimized skills are then uploaded into the AI system, enabling all Atlas robots to acquire new capabilities.
Scott Kuindersma, head of robotics research at Boston Dynamics, attributes the robot's ability to learn movements previously considered challenging for machines to this shift towards teaching, demonstrations, and machine learning over manual programming.
Current Limitations and Future Outlook
While Atlas has demonstrated advanced capabilities, limitations persist. Kuindersma noted that Atlas is not yet proficient at routine daily tasks that humans perform, such as dressing or pouring coffee, though a pathway to achieving such capabilities is becoming evident. Challenges also remain in refining teleoperation systems, particularly concerning the precise control of gripper shape, motion, and force for more dexterous manipulation tasks.
Robert Playter, CEO of Boston Dynamics, estimates it will be several years before Atlas becomes a full-time worker. He stated that while advancements in AI and software can occur rapidly, the development and deployment of reliable and affordable robotic hardware require a considerable timeline. Playter also indicated that humanoids could alter the nature of work, with robots undertaking repetitive and physically demanding labor, while still requiring human management, construction, training, and servicing. Playter highlighted the potential for robots like Atlas to offer enhanced strength, heat tolerance, and the ability to operate in hazardous environments.
Market Projections and Societal Implications
Goldman Sachs projects the humanoid robot market to reach $38 billion within the next decade. U.S. robot manufacturers, including Boston Dynamics, operate within a competitive global market that includes state-supported Chinese companies. Playter acknowledged the technical lead of U.S. firms but also identified a potential challenge from the scale of investment by competitors.
Concerns exist regarding potential worker displacement by AI, as Atlas is being trained for tasks currently performed by human workers at Hyundai's Georgia plant. Industry forecasts project a future with substantial integration of robots into daily life.