The evolution of the Atlas humanoid robot at Boston Dynamics has marked a significant shift in focus, moving beyond mere mobility and locomotion to concentrate intently on sophisticated manipulation capabilities. As explained by Alberto Rodriguez, this transition was naturally prompted by the robot's move from hydraulic systems to an electric platform, initiating an intense exploration into high-dexterity grippers. Engineers recognized from the outset that the development of robotic hands is a protracted journey, given that grippers are arguably the most intricate components of a humanoid, requiring vast functionality, actuation, and sensing to be packed into an extremely confined space—a challenge described as a "very hard design problem". Boston Dynamics has adopted a long-term perspective to maximize learning through the development of what they term the "GR line". The initial effort, GR1, served to define the most fundamental, minimalistic capabilities of a human hand. Crucially, GR1 taught the team about ruggedness, a vital characteristic for a development path where robots inevitably fall, sometimes directly onto the gripper.
Related article - Uphorial Radio
The lessons learned paved the way for the second-generation gripper, GR2, which represents a significant technological leap. GR2 features seven degrees of freedom and houses seven distinct actuators—two for each of the three fingers and one dedicated to the articulated thumb joint. The entire module is self-contained and designed for easy removal and attachment. Perhaps the single most important innovation distinguishing GR2 from its predecessor is the addition of an opposable thumb. This inclusion dramatically expands the flexibility of potential grasps, allowing the robot to manipulate "almost anything that we throw at it". The design settled on three fingers because engineers believe this is the fewest number necessary to successfully execute highly complex manipulation tasks. The geometry of three fingers, utilizing distance between the digits, is essential for maintaining a stable grasp, particularly when handling objects that are larger, heavier, or prone to rotation. While the opposable thumb allows for delicate two-finger pinch grasps needed for handling tiny objects, the third finger is instrumental in ensuring stability when heavier items are involved. Although engineers debated and considered adding more fingers, the decision was made to halt at three, recognizing that additional complexity, unless strictly necessary, typically results in lower reliability, higher cost, and slower development speed. However, this three-finger approach is currently considered the correct answer for the program’s initial phases, and not necessarily a permanent "dogma".
Core to the gripper’s sensitivity is its advanced tactile sensing system, which functions as force feedback—analogous to a human’s sense of touch. These sensors are embedded in the fingertips, situated beneath an elastomer material that serves a dual purpose: providing a high-friction surface and deforming under pressure, allowing internal sensors to translate that deformation into force feedback. This sense of touch is critical for applying gentle forces when manipulating fragile objects and also allows the robot to detect if an object has been dropped. The primary objective is to apply the minimum necessary force to maintain a stable grasp, avoiding either crushing the object with too heavy a grasp or allowing it to slip away due to insufficient pressure. In terms of kinematics, the gripper can rotate up to 90 degrees, similar to a human wrist, but possesses an extra capability: the fingers can bend completely backward, enabling the robot to grasp objects on the back side of the unit. The design features distinct left and right versions, mirrored like human hands. Unlike humans, however, Atlas does not exhibit a dominant hand; it strategically plans the most optimal route for a task, choosing the left or right hand based on stability requirements or the need to avoid environmental obstacles. Looking forward, the next critical milestone is increased dexterity, driven by the prevalence of real-world tasks in manufacturing such as bin picking, tool use, and the handling of small components. These practical requirements are organically pushing the design toward more anthropomorphic forms. The ongoing challenge—and one of the most exciting journeys for the coming years—remains finding the optimal "sweet spot" that balances dexterity, actuation, and sensing capabilities.
