Boston Dynamics unveils a new four-fingered hand for the Atlas humanoid robot

Boston Dynamics has unveiled a new four-fingered hand for its Atlas humanoid robot, designed to combine the dexterity required for complex manipulation and tool use with the robustness demanded by industrial work. According to the company, it replaces the previous seven-degree-of-freedom hand, which was conceived primarily for grasping objects, with a design aimed specifically at manipulating them.

The new hand has 13 degrees of freedom in total: each of the three fingers contributes three degrees of freedom, while the opposable thumb accounts for four. That thumb can slide both along and across the other fingers, allowing Atlas to adopt different pinch and tool-holding configurations. The hand is comparable in size to a large human hand, in part because Atlas is intended to operate in spaces, environments and workstations originally designed for people.

Boston Dynamics explains that, despite the greater dexterity, the new hand retains strength similar to that of the previous design; the company notes that it has already shown Atlas carrying a mini-fridge weighing more than 100 pounds (around 45 kilos). Among the capabilities the manufacturer highlights are in-hand object reorientation, pinch and tripod grasps, recovery when an object slips, and the handling of tools such as drills, powered torque screwdrivers, grinders, nail guns and welding torches. The actuators are encapsulated, the design is based on a single actuator type and no cables cross the joints, an approach that, according to the company, improves durability, manufacturability at scale and ease of repair with a view to larger-scale humanoid robot production.

One striking aspect of the announcement is the rationale for why the hand has four fingers rather than five. Boston Dynamics notes that, during development, the company's chief technology officer, Zack Jackowski, asked members of his team to tape their pinky and ring fingers together for a day and write down which tasks they could no longer perform. That experiment helped convince the team that four fingers were enough for the capabilities they were pursuing, whereas adding a fifth finger would have required three additional actuators, increasing cost, volume and the number of potential failure points. With that four-fingered architecture, Atlas can still reorient objects in the hand, recover from slipping grasps and hold tools while pulling their triggers.

The hand also incorporates densely distributed tactile pressure sensors on the fingertips and palm, which add to the proprioceptive sensitivity of the actuators themselves. According to Boston Dynamics, this makes it possible to detect very small contact signals during interaction with objects and provides the feedback needed for more delicate manipulation. Backdrivability and actuator transparency are also central design elements, since they allow proprioception to be used for agile manipulation and precise force regulation.

Tool use is, in fact, one of the project's stated focuses: rather than trying to create a hand capable of performing any task directly, Boston Dynamics sees tools as a way to extend the capabilities of humanoid robots, in the same way tools extend human capabilities. The company argues that this will become increasingly relevant as humanoids move beyond conventional pick-and-place operations to tackle assembly tasks, cable and tool handling, and objects presented in less structured environments.

The design was also conceived with simulation and reinforcement learning in mind. Boston Dynamics says the hand's stiff-drive actuation, backdrivable transmission and control systems make it possible to reproduce its physical behavior with high fidelity in simulation, which makes it easier to train manipulation policies by varying factors such as torque, surface friction, object geometry and different disturbances, before transferring those behaviors to the physical robot. The company says it has already achieved initial sim-to-real transfer results on dynamic manipulation tasks, with behaviors trained entirely in simulation and then deployed on Atlas hardware, relying on the proprioception of the high-frequency actuators as feedback. Boston Dynamics views reinforcement learning as a complement to human demonstration data: the latter would provide a large amount of information about how objects are manipulated, while reinforcement learning in simulation would train the high-speed control and force regulation needed to execute those movements reliably on the robot.

This hand is presented as a complement to the latest generation of Atlas, the fully electric humanoid robot Boston Dynamics is developing with industrial applications and eventual large-scale manufacturing in mind. The company maintains that dexterous manipulation with tools will be a core capability if humanoid robots are to become flexible machines capable of performing a wide range of real industrial tasks.