Standard robot arms are intrinsically stiff due to their linkage designs and gear ratios, typically greater than 80:1, which provide great strength and precision. However, this gearing amplifies the robot’s effective inertia and mass, making compliant behavior impossible. When the robot makes contact with an object, the inertia created by the high gear ratios can cause the robot to apply undesirable or damaging forces. For a robot to complete tasks that require strength and tasks that require compliance, it needs both the ability to produce high forces but also the ability to control them during contact.

The RAI Institute is building robot manipulators that are designed for compliant contact with lower reflected inertias at its joints.  The goal is to make a manipulator that is fast and precise, and that can switch seamlessly between delicate and forceful contact.  AthenaZero, our first robotic manipulator, was designed to be both dynamic and agile. It can be gentle but also apply high forces when necessary. Our design achieves this by paying attention to the effective mass properties – how much the robot arms weigh, where that weight is concentrated, and the gear ratio and linkage that determine how easy it is for the motors to be driven backwards. The mass properties of AthenaZero are comparable to human arms, approximately an order of magnitude lower than leading collaborative research robots with ratios of <11:1 and 5:1 for most joints.

Challenges with Dynamic Manipulation at a Human Cadence

In addition to keeping the effective mass low, we chose human cadence, or the tempo in which a human completes a particular task, as a performance goal.  Robot manipulation has achieved human cadence while doing dexterous tasks only for a few specific tasks, but not on a broad set of tasks. Our thesis is that most commercial manipulators are not designed to make achieving such behavior easy—their  joints are natively too stiff for the dynamic changes and fast accelerations seen in much human behavior. 

To achieve the goal of human cadence performance and compliant performance, we use the  inertia of the robot  as an upper bound on how compliant a particular robot arm design can be and as a parameter  for optimization. Higher inertia equates to a robot that has a harder time accelerating (required to match human cadence) and decelerating (required to achieve compliance). Control systems, both modern and classical, are limited by the need to compensate for high inertia, thus forcing slowed down behavior to achieve delicate or gentle performance.

The key metric is the effective mass of the robot at the  point of environmental contact. If we can reduce the effective mass at  points of contact, the acceleration/deceleration at that contact interaction can be performed in a more fluid, human-like manner. The largest contributor to effective mass is usually actuator-reflected inertia, how much resistance to motion a motor “feels” from the load it’s driving, which changes depending on the gear ratio between the end effector and the shoulder or base of the robot. Using stronger motors with low gear ratios can substantially reduce actuator reflected inertia and can move manipulator characteristics closer to those of a human arm.

A Balancing Act: Maintaining High Control Authority While Minimizing Manipulator Inertia

AthenaZero is a bimanual robot prototype consisting of a one degree of freedom (DoF) torso, two 7-DoF arms, and two 6-DoF underactuated hands (22 actuators over a total of 27 joints). The robot stands approximately 5’3” tall (1.6 m) with a wingspan of roughly 6 ft (1.8 m) The platform’s core design philosophy prioritizes maximizing control authority – or the ability to accelerate quickly – and minimizing effective mass, while maintaining support for payloads greater than 6.6 lb (3 kg). This combination of characteristics can, we believe, enable the platform to achieve human-like dynamic performance in a wide range of manipulation tasks. The key concept is that, much like humans, robots should be designed to strategically leverage contact rather than avoid it, and the first step in achieving that is building robots that exhibit the qualities of control and inertia.

To accomplish such performance, the system leverages two main techniques: quasi-direct drive actuators with efficient planetary gearsets and shifting much of the mass closer to the robot’s center of mass. A family of four custom low gear ratio actuators in different diameter variations were designed and utilized. As the actuator’s reflected inertia, a key ingredient of the effective mass, scales by the square of the gear ratio, we kept these values low, using 5:1 in most joints.  We also made use of efficient planetary gearsets (97% or better). This combination allowed us to rely entirely on motor current measurements for torque sensing. Unlike typical arms with force sensors, there are no force-torque sensors anywhere on the robot, thus eliminating a fragile sensor and a source of noise (through drift). Another key feature of the robot that results from this design is backdrivability, which means the system can transmit force back through the geartrain. Unlike traditional, commercial robots that are often dominated by friction in numerous gear stages, this design allows the system to comply with external forces, if desired.  The low gear ratio also allows energy provided by proximal actuators to power motions of distal joints as in the wind-up and pitch of a baseball player.

The second method to reduce manipulator inertia is to move actuation more proximal to the robot’s base, its torso. This optimization reduces the center of mass distance from the base to the robot links as the heavy components of the robot (i.e., the motors) are closer to the torso of the robot, thereby reducing the amount of moving mass when swinging the arm through space.

Baseball Tasks as a Testbed

We chose to test the characteristics of AthenaZero using the athletic game of baseball. We selected baseball not because we hope to send a robot onto the baseball field someday but because baseball requires both precision and agility.

Three baseball-inspired tasks were used to validate the robot’s dynamic capabilities: throwing, catching, and batting.

  • Throwing: Through dynamic trajectory optimization, AthenaZero threw a ball 70 mph (113 kmh). This speed is approaching human amateur ability as a typical high school pitcher throws between 70-80+ mph (113-129 kmh), and a college pitcher between 80-90+mph (129-145 kmh). 
  • Catching: Through impedance matching, the robot was able to catch a ball at up to 41 mph (66 kmh) over a short distance of 24 ft (7.3 m). Extrapolated to 60ft 6in (18 m) to match a true pitcher’s mound, and keeping reaction time constant (0.398s), this performance is like catching a 103 mph (166 kmh) fastball (professional pitcher speeds).
  • Batting: Over 33 batting attempts (swings) AthenaZero made contact at 31 mph (50 kmh) 82% of the time. Similarly extrapolated to the distance of a true pitcher’s mound, this is equivalent to batting at about 78 mph (126 kmh), solidly placing it at an amateur level. 

The system’s capabilities were further tested through robot-to-robot and human-to-robot variations of playing catch and batting practice. In robot-to-robot catch at approximately 30 mph (48 kmh), the two robots volleyed successfully up to 8 times with 11 second catch-to-throw cycles. When one robot was replaced with a human participant, the system successfully volleyed up to 12 times, showing adaptability to trajectory uncertainty. Batting practice operated for approximately 3 minutes in both configurations, with the robot adapting its swing trajectory in real-time based on continuous estimation of the ball’s projected position.

Comparative Analysis: How AthenaZero Stacks Up

Comparative analysis against leading collaborative robots confirms significant improvements in effective mass properties. At the wrist and in a neutral joint configuration, AthenaZero achieves effective mass of 9 lb (3.97 kg) which is comparable to a human arm at 6 lb (2.76 kg) and far less than the Franka Research 3 at 64 lb (29.21 kg) and the Universal Robots UR5e at 77 lb (34.72 kg). Although AthenaZero’s post-shoulder mass is nearly twice that of a human arm, placing actuators closer to the shoulder reduces distal mass. For commercial robots, reflected inertia scales quadratically with the actuator’s gear ratio, which quickly increases the effective mass. AthenaZero’s effective mass is minimally impacted by actuator contributions, with inertia dominated by links rather than actuators.

The advantage becomes especially clear in tasks involving contact along the entire arm. Heat map analysis shows that AthenaZero maintains relatively low effective mass even for contacts on the lower and upper arm, with upper arm values comparable to those on the distal links of commercial robots. This property makes the system well-suited for whole-body manipulation involving contact beyond the hands, functionality that we hope to demonstrate soon.

Technical Considerations and Limitations

AthenaZero’s low-inertia design requires control approaches that differ from those used with traditional manipulators. The system’s dynamical behavior requires careful attention to torque control and trajectory acceleration components for feedforward dynamics that go well beyond kinematic planning or feedforward position control. Our research shows that the synthesis of human-looking, graceful motions in robot manipulation is not simply a product of minimal jerk optimization, which may contradict common musings in both robotics and biomechanics. Multiple studies since the 1980s have shown that minimum jerk does not explain human motion in general.

For example, the mechanical principle in throwing is the so-called kinetic chain, where the motion of a limb starts when the previous limb reaches maximum speed. In the case of catching, there are competing explanations (about what humans leverage). We implemented an approach based on task readiness impedance, to mirror how humans regulate their hand impedance properties appropriately in target tasks that involve dynamic contact.

The same features that make AthenaZero suitable for leveraging contact and producing agile movements also make it less suitable for certain traditional manipulation tasks. Like humans, AthenaZero is not well-suited for sustained torque output over long durations due to thermal dissipation, such as static holds of an object distal to the torso for over a minute, or tasks demanding high endpoint stiffness, such as  strict trajectory following, like in welding. However, the transparency and backdrivability of the system make it more tolerant to errors in the environment. We find that forceful assembly tasks actually benefit from robot compliance (i.e., flexibility during contact). This property opens the door to the development of more flexible, human-like control strategies that may ultimately prove more effective for dynamic, contact-intensive manufacturing and assembly tasks – an exciting paradigm shift in how we think about controlling robots.

From Baseball Player to General-Purpose Manipulator

AthenaZero demonstrates that contemporary motor technology allows us to get manipulator designs that are closer to human-like dynamic properties, moving towards the goal of a general-purpose manipulator with capabilities similar to that of a human. By prioritizing and leveraging effective mass reduction and force transparency through quasi-direct drive actuators and strategic positioning of the mass, the system achieves performance levels previously unattainable in robot arms. 

Approaching the flexibility of human-level manipulation requires robotic systems that work with contact dynamics, not against them. In the future, we will focus on natural interaction performance in unpredictable environments, including the development of further appropriate control strategies and the study of the factors behind natural dynamic behavior.